Introducing DART: A Novel Deep Adaptive Upsampling Technique for Handling Class Imbalance

Introducing DART: A Novel Deep Adaptive Upsampling Technique for Handling Class Imbalance | Research Square window.SnipcartSettings = { analytics: { enabled: false } }; (function() { var accessVector = localStorage.getItem('access_vector') || ''; window.dataLayer = window.dataLayer || []; if (accessVector) { window.dataLayer.push({ user: { profile: { profileInfo: { snid: accessVector } } } }); } })(); (function(w,d,s,l,i){w[l]=w[l]||[];w[l].push({'gtm.start':new Date().getTime(),event:'gtm.js'});var f=d.getElementsByTagName(s)[0],j=d.createElement(s),dl=l!='dataLayer'?'&l='+l:'';j.async=true;j.src='https://www.googletagmanager.com/gtm.js?id='+i+dl;f.parentNode.insertBefore(j,f);})(window,document,'script','dataLayer','GTM-K279D39R'); Browse Preprints In Review Journals COVID-19 Preprints AJE Video Bytes Research Tools Research Promotion AJE Professional Editing AJE Rubriq About Preprint Platform In Review Editorial Policies Our Team Advisory Board Help Center Sign In Submit a Preprint Cite Share Download PDF Research Article Introducing DART: A Novel Deep Adaptive Upsampling Technique for Handling Class Imbalance Mark Lokanan This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-6895500/v1 This work is licensed under a CC BY 4.0 License Status: Posted Version 1 posted You are reading this latest preprint version Abstract Class imbalance remains a persistent challenge in predictive modeling, often leading to biased machine learning outcomes that disproportionately favor the majority class. This study investigates the effectiveness of advanced resampling techniques—both undersampling and oversampling—across two large and highly imbalanced datasets involving credit and loan default prediction. In addition to evaluating established oversampling techniques, the study introduces and validates a novel resampling approach, DART (Deep Adaptive Resampling Technique). Each technique is assessed using a consistent suite of classifiers, including logistic regression, gradient descent, Naïve Bayes, random forest, CatBoost, and artificial neural networks. The results reveal that K-MeansSMOTE and NearMiss outperform other resampling strategies in oversampling and undersampling, respectively, by achieving balanced trade-offs in precision, recall, F1-score, AUC, and Matthews Correlation Coefficient. Notably, DART demonstrates exceptional performance across both datasets, achieving nearly perfect classification scores across all metrics, suggesting strong generalizability and robustness. The study further analyzes the strengths and limitations of each resampling technique and emphasizes the importance of metric selection when evaluating imbalanced datasets. By integrating empirical evaluation with theoretical insights, this research contributes to the growing body of literature on imbalanced learning and offers practical guidance for selecting appropriate resampling strategies. These findings have broader implications for domains such as finance, healthcare, and fraud detection, where class imbalance is common. Overall, the study affirms the value of hybrid and adaptive resampling methods in building more accurate and generalizable predictive models. Artificial Intelligence and Machine Learning Undersampling Oversampling Machine learning Credit Default Loan Default DART Imbalance dataset Figures Figure 1 Figure 2 Figure 3 Figure 4 1 Introduction The success and accuracy of machine learning models depend heavily on the quality and distribution of the data. One of the most persistent challenges in this regard is class imbalance, which can significantly distort model performance by biasing predictions toward the majority class. Imbalanced data occurs when one class is heavily underrepresented compared to the other, leading to biased predictions and poor generalization, especially for the minority class. Data imbalance is frequently encountered in real-world scenarios such as credit and loan default prediction, fraud detection, and disease diagnosis, where the event of interest (e.g., default, fraud, or illness) occurs far less frequently than the non-event. Imbalanced datasets present numerous challenges for machine learning algorithms (Araf et al., 2024 ; Cinaroglu, 2020 ). The skewed distribution of classes can introduce biases during model training, favouring the majority class and resulting in diminished performance for the minority class. The dominance of the majority class within the dataset often leads algorithms to prioritize predicting this class, potentially overlooking the minority class, resulting in biased classifications and poor generalization. Imbalanced classification scenarios, also known as rare event modelling, arise when the target variable exhibits a significant imbalance, with the minority class representing the rare events. In such cases, the model's tendency to learn predominantly from the majority class can make predicting the minority class particularly difficult (Araf et al., 2024 ; Khushi et al., 2021 ; Naseriparsa & Kashani, 2014 ). Consequently, machine learning algorithms may struggle to construct accurate models (Chawla et al., 2002 ), leading to challenges in evaluation metrics that could fall into the "metric trap" and result in inaccurate results (Jeni et al., 2013 ; Yap et al., 2014 ). When handling imbalanced data, many machine learning algorithms tend to excel at accurately predicting the majority class while struggling with the minority class (Zhou, 2013 ). In datasets with class imbalances, classification algorithms often prioritize predicting the majority class without adequately analyzing the minority class (T. Sun & Vasarhelyi, 2018 ). Consequently, this can lead to consistently high-performance accuracy scores that are not necessarily reliable (An & Suh, 2020 ; Hooda et al., 2018 ; Lin et al., 2023 ). Relying solely on raw performance accuracy may not be ideal for evaluating classification models, especially in the context of imbalanced datasets (Araf et al., 2024 ; Hooda et al., 2018 ). This outcome is because accuracy only considers true positive/negative observations and ignores false positive and negative observations. While alternative metrics such as precision, recall, F1-score, and Area Under Receiver Operating Curve (ROC) Curve (AUROC) offer more robust evaluation options for imbalanced datasets, they still tend to favour the majority class (Fujiwara et al., 2020 ; Hasanin et al., 2019 ). The purpose of this study is to tackle class imbalance in predictive modeling, a problem that can severely distort machine learning performance by favoring the majority class. Rather than focusing on general algorithmic adjustments or cost-sensitive learning, this paper concentrates on resampling strategies—both undersampling and oversampling—as effective methods for correcting skewed class distributions. To achieve this purpose, the study is guided by two main objectives: To systematically evaluate and compare the effectiveness of established resampling techniques across a diverse set of machine learning classifiers using both a credit and a loan default dataset. To introduce and assess the performance of a novel resampling method, DART (Deep Adaptive Resampling Technique), by applying it to both the credit default dataset and a secondary loan default dataset, thereby testing its robustness and generalizability. In addressing these objectives, the paper makes several important contributions to the literature on handling imbalanced datasets. First, while prior studies often focus on a single undersampling or oversampling method applied to one dataset, the present study conducts a comprehensive evaluation of multiple resampling techniques across two distinct datasets—credit default and loan default. By assessing the performance of each method across a range of machine learning algorithms and datasets, the study provides a broader and more robust understanding of which techniques yield the most effective results in different predictive contexts. Second, the comparative analysis of undersampling and oversampling methods generates practical scientific insights that bridge the gap between theoretical advancements and real-world applications. Highlighting the trade-offs, strengths, and limitations of each technique empowers researchers and practitioners to make more informed methodological choices when working with imbalanced data. Third, a novel resampling technique—DART—is introduced and applied to both datasets. DART dynamically adjusts the sampling process based on model feedback, offering a flexible and adaptive solution to the class imbalance problem. Its integration into the modeling pipeline demonstrates its potential to improve classification performance, particularly in the domain of financial risk prediction. The structure of this paper is organized as follows. Section 2 presents a review of the literature on undersampling techniques, highlighting their theoretical foundations and applications. Section 3 examines the body of research on oversampling methods, with particular emphasis on SMOTE-based and hybrid approaches. Section 4 critically evaluates the advantages and drawbacks of both undersampling and oversampling strategies, especially in the context of model performance on imbalanced datasets. Section 5 outlines the methodology, detailing the datasets, resampling techniques, and classification algorithms employed. Section 6 provides a comprehensive analysis of the empirical results, comparing the effectiveness of the various resampling approaches across multiple metrics. Finally, Section 7 offers a discussion and conclusion, synthesizing the key findings and outlining their implications for future research and practical applications. 2 Literature Review: Resampling Imbalanced Datasets A well-established method for addressing the class imbalance problem is resampling the data (Chawla et al., 2002 ; Ghorbani & Ghousi, 2020 ; Lin et al., 2023 ). The objective of resampling is to ensure that the samples used in the model closely resemble the population they originate from, facilitating accurate inferences about the true population from the sample (Rajer-Kanduč et al., 2003 ). Resampling encompasses two main approaches: undersampling, which involves removing observations from the majority class to align with the minority class, and oversampling, which entails adding more observations to the minority class (An & Suh, 2020 ; Cinaroglu, 2020 ; Xuan et al., 2018 ). Both undersampling and oversampling techniques aim to adjust the class ratio in an imbalanced dataset for modelling purposes. The subsequent section reviews the literature on the undersampling and oversampling techniques that have gained prominence over the years. Special attention will be given to their effectiveness in addressing class imbalances, their application in various domains, and their impact on model performance and generalization. The literature to follow is illustrative rather than exhaustive. The intention is to explore the use and applicability of undersampling and overs-sampling techniques in machine learning applications that have been in ascendency in the literature. 2.1 Undersampling Majority Class Undersampling is a technique where the number of observations from the majority class is reduced to match the minority class (An & Suh, 2020 ; Dal Pozzolo et al., 2014 ; Zuech et al., 2021 ). Figure 1 presents a diagrammatic illustration of the undersampling technique, where there are 12,000 observations of the majority class and 2,000 observations of the minority class. A model built with this sample distribution will be biased towards the majority class because the sample is more likely to be fed into the algorithm multiple times compared to the minority class (An & Suh, 2020 ; Hernandez et al., 2013 ; Santos et al., 2018 ). As shown in Fig. 1 , to create a balanced dataset, the data is resampled with random samples of 2,000 observations from the majority class to match the 2,000 observations of the minority class (An & Suh, 2020 ). Undersampling is often favoured due to the presence of a sufficient minority class, allowing for a balanced 50–50 split while maintaining a representative sample. This approach aids data mining algorithms in effectively analyzing the data within manageable limits (Dal Pozzolo et al., 2014 ; Fujiwara et al., 2020 ). With undersampling, the majority of class observations are discarded at random until there is a more balanced distribution of the data. Balancing out the minority and majority classes allows the classifier to weigh both classes equally and produce more representative results (Dal Pozzolo et al., 2014 ; Hernandez et al., 2013 ; Yap et al., 2014 ). Several undersampling techniques commonly utilized in the literature include random undersampling, Tomek links, NearMiss, and Condensed Nearest Neighbour (CNN). Below, we examine the existing literature assembled to date on these undersampling techniques. 2.2 Random undersampling Random undersampling (RUS) is a technique used in machine learning to address the issue of imbalanced datasets, where the majority class significantly outnumbers the minority class (Kamei et al., 2007 ; Saripuddin et al., 2021 ; Zuech et al., 2021 ). This approach involves randomly selecting a subset of samples from the majority class and removing them from the dataset to achieve a more balanced class distribution. Practitioners can implement RUS by specifying either a desired ratio of majority to minority class samples or a fixed number of samples to be removed from the majority class. RUS is simple to implement and computationally time- and resource-efficient (Zuech et al., 2021 ). Recent studies on RUS in machine learning span across various domains, each exploring the technique's effectiveness in addressing class imbalance problems in specific domains. Some researchers have employed RUS to improve classification performance in anomaly detection (Huan et al., 2020 ; Saripuddin et al., 2021 ; Y. Yang et al., 2023 ). These studies underscore the versatility of RUS in different domains to detect anomalies and improve the performance of machine learning models. Others have utilized RUS to address the challenges of class imbalance in cybersecurity datasets (Bagui & Li, 2021 ; Silva et al., 2021 ; Zuech et al., 2021 ). By employing RUS, these researchers were able to enhance detection accuracy for cybersecurity threats by making the models more reliable and efficient in identifying potential threats in vast amounts of normal traffic data. Another area where RUS has been utilized with very good performance is in predictive modeling and disease diagnosis in the healthcare industry (Bauder & Khoshgoftaar, 2018 ). Researchers have applied RUS to Medicare fraud detection, showing that it significantly boosts classification accuracy (J. Hancock et al., 2022 ). Pias et al. ( 2023 ) explore the use of RUS to balance datasets in predicting diabetes and prediabetes in patients. The authors found that RUS achieved robust results and enhanced the fairness and performance of machine learning models (Pias et al., 2023 ). Others have noted that a 50:50 RUS does not produce the best Medicare fraud detection results; rather, the authors found that a 90:10 class distribution offers the best detection (Bauder & Khoshgoftaar, 2018 ). These findings suggest that RUS may not be the best technique for fraud prediction in the healthcare industry. That aside, these studies collectively underscore the significance of RUS in enhancing the performance of machine learning models in a variety of domains. 2.3 Tomek links The Tomek links undersampling method is used to balance imbalanced datasets by removing observations located between two different classes. The basic idea behind Tomek links is to identify pairs of observations, one from the majority class and one from the minority class, that are closest to each other but belong to different classes (Alamri & Ykhlef, 2024 ; Devi et al., 2017 ). These pairs are called Tomek links (Devi et al., 2017 , p. 3). Tomek links works by eliminating instances belonging to the majority class from Tomek linkages. Removing the majority class can effectively enhance the separation between classes and boost the efficiency of classification systems. Tomek links focuses on the ambiguous cases that are prone to causing misclassification and eliminates them to provide a more distinct decision boundary between the classes. Tomek links aims to improve the overall balance of the dataset while preserving the minority class instances that are farthest from the majority class. Research in Tomek links to address class imbalance spans a variety of domains and demonstrates the versatility and effectiveness of the technique to enhance the performance of machine learning models in different applications. Some studies propose an approach combining one-class SVM for anomaly detection with adapted Tomek links pairs to eliminate redundant and overlapping cases and address imbalance data issues (Basit et al., 2022 ; Devi et al., 2017 ; Vuttipittayamongkol et al., 2021 ). Interestingly, the findings suggest that class overlap has a more detrimental effect on performance than class imbalance. That said, there is evidence to show that Tomek links has been very effective in addressing imbalances and overlapping issues in medical datasets (Basit et al., 2022 ). Others have used Tomek links in the area of multi-label classification to remove boundary and noise samples from datasets (Ai-jun & Peng, 2020 ; Pereira et al., 2020 ). Like Devi et al. (2019), these studies address imbalance by selectively removing overlapping samples between classes to improve the performance of multi-label classification (Ai-jun & Peng, 2020 ; Pereira et al., 2020 ). Others have used Tomek links in bioinformatics and medical diagnostics to improve machine learning model performance on imbalanced datasets (Ning et al., 2022 ; Zeng et al., 2016 ). The authors note that Tomek links, in combination with the synthetic minority oversampling technique (SMOTE), enhances classification accuracy across different metrics and demonstrates the benefits of combining these resampling techniques for medical data classification (Zeng et al., 2016 ). These studies illustrate a growing interest in using Tomek links to address the intertwined challenges of class imbalance and overlapping issues across different applications. 2.4 NearMiss NearMiss is an undersampling method that balances imbalanced datasets by removing observations from the majority class closest to the minority class. NearMiss is a k-nearest neighbor approach that balances the class distribution by choosing instances based on the distance between the majority class and the minority class (Mqadi et al., 2021 , pp. 3–4; Oladunni et al., 2021 , p. 3). When the distance between the majority and minority classes is too close, NearMiss removes instances of the majority class in order to increase the distance between them (Ha & Lee, 2016 , p. 2). To find the closest instances of the majority class, there are three types of NearMiss techniques: NearMiss-1, NearMiss-2, and NearMiss-3. NearMiss-1 selects instances from the majority class whose individual distance to the three closest instances to the minority class is the smallest; NearMiss-2 selects instances from the majority class whose individual distance to the three farthest instances to the minority class is the smallest; and NearMiss-3 selects instances from the majority class for which each instance in the minority class has the closest distance (Ha & Lee, 2016 , p. 2). The commonality of the NearMiss family is that they select the instances in the majority class that are closest to the minority class to better learn the decision boundary in the data (p. 2). Most of the literature on NearMiss undersampling is sparse, with few studies exploring its effectiveness across various domains. Bao et al. ( 2016 ) proposed boosted near-miss undersampling on SVM (BNU-SVMs) ensembles for concept detection in large-scale imbalanced datasets. The authors discovered that BNU-SVMs may effectively manage large-scale imbalanced datasets by balancing and reducing the training dataset through undersampling (L. Bao et al., 2016 ). The classifier's performance is enhanced by integrating multiple classifiers. Other studies have employed NearMiss undersampling to address imbalances in financial crime datasets (Mqadi et al., 2021 ; Rubaidi et al., 2022 ). The studies found that machine learning algorithms performed very well using the NearMiss undersampling technique. NearMiss undersampling has also been used in detecting insider threats, specifically data leakage by malicious insiders prior to leaving an organization (Alsowail, 2022 ). The author found that NearMiss undersampling achieved enhanced performance in improving the detection of insider data leakage. Further research in healthcare has found NearMiss to be a promising undersampling technique in bioinformatics and predicting disease with high performance accuracy (Alamsyah et al., 2022 ; Nayan et al., 2023 ). 2.5 Condensed Nearest Neighbor (CNN) The Condensed Nearest Neighbor (CNN) undersampling algorithm operates by systematically removing redundant instances from the majority class that are correctly classified by their nearest neighbor in the minority class while preserving essential information in the dataset. The algorithm follows two main steps: Initially, a random subset of the majority class samples is selected to form the "condensed set." The algorithm then iteratively eliminates samples from the majority class that the condensed set's nearest neighbor correctly classified. This procedure keeps iterating until every sample in the condensed set receives the wrong classification from their nearest neighbor in the majority class. The CNN algorithm employs the 1-NN rule, which dictates that all minority class instances are assigned to set S, one majority class instance is placed in set S, and the remaining majority class instances are allocated to the set (Chaplot et al., 2019 , p. 95). Each sample from set C is individually assessed using the 1-NN rule (p. 95). If correctly classified, the sample is discarded; otherwise, it is moved to set S (p. 95). This process repeats for all instances until all the instances have been evaluated using the 1-NN rule (p. 95). The underlying concept behind CNN is that instances correctly classified by their nearest neighbor are considered uninformative for establishing the decision boundary between classes and can be safely removed without compromising classification performance (Batista et al., 2004 ; Chaplot et al., 2019 ; Devi et al., 2017 ). While CNN is limited in application, there are a few studies that provide evidence of its effectiveness in addressing class imbalance. In their study, Bansal and Jain employed CNN as a method to balance the number of instances between two classes (Bansal & Jain, 2021 ). They achieve this by undersampling the majority class according to specific criteria and finding that CNN was the best performer among others. Batista et al. tested CNN along with other techniques across thirteen different datasets and found that class imbalance does not hinder the performance of learning systems (Batista et al., 2004 ). The problems appear to be rooted in learning with too few minority classes in the presence of other heterogeneous factors such as class overlapping (p. 20). In another study, Xie et al. 2021 employed CNN and other undersampling techniques on 40 public benchmark datasets. The study discovered that CNNs eliminate noisy or boundary occurrences from the majority class, which can be advantageous for learning models dealing with imbalanced data (Xie et al., 2021 ). Nevertheless, even when using undersampling techniques, a significant disparity between the majority class and the minority class might still negatively impact the performance of the learning process (Xie et al., 2021 , p. 8). Each undersampling strategy has its own distinct advantages in mitigating class imbalances in machine learning applications. RUS has exhibited adaptability and enhanced effectiveness in multiple areas, such as anomaly detection, cybersecurity, and healthcare. Its implementation has been recognized for its ability to improve machine learning models. However, Tomek links specifically aims to address ambiguous observations between different classes. Tomek links has shown potential in the fields of medical diagnostics and multi-label classification, where it has been successfully utilized to enhance model performance by getting rid of irrelevant information and resolving overlaps in the data. NearMiss effectively enhances the distinction between classes by deliberately eliminating instances of the majority class that are closest to the minority class. NearMiss has been employed to enhance machine learning models in the domains of financial crime detection and healthcare. Although CNN and other undersampling approaches provide potential for addressing class imbalance problems in machine learning, the continued existence of substantial differences between the majority and minority classes highlights the persistent difficulty in reaching optimal performance across varied datasets. While these undersampling strategies offer unique advantages in mitigating class imbalance across diverse domains, challenges persist in achieving optimal performance due to significant disparities between the majority and minority classes. Oversampling techniques provide an opportunity to obtain insights on how these methods have been used to address class imbalances in datasets. 3 Oversampling Minority Class Random oversampling of the minority class is done by duplicating the instances and their representation in the dataset (More, 2016 , pp. 5–8). Referring back to the fraud (2%) and non-fraud (98%) examples above, the majority of the data will be skewed towards the no-fraud class. In this case, because there is only 2% of the fraud class, the model will mostly train on the 98% of the no-fraud class. One way of oversampling is to generate new instances for the minority class by sampling with replacement (An & Suh, 2020 ; Chawla et al., 2002 ). Figure 2 presents a diagrammatic illustration of random oversampling. Note from Fig. 2 that 2000 transactions were duplicated six times to balance the data. Oversampling is preferred when there is an abundance of data for the majority class and rare events for the minority class (Elreedy & Atiya, 2019 ; Yap et al., 2014 ). 3.1 Random Oversampling Random oversampling (ROS) is a method used in handling imbalanced datasets, where instances from the minority class are duplicated randomly to augment the dataset until it reaches the desired ratio or balance with the majority class (Elreedy & Atiya, 2019 ; Nayan et al., 2023 ). Unlike other techniques, ROS does not consider the similarity and characteristics of the data points and simply duplicates the instances without considering the relevant features in the dataset. ROS typically involves the following steps: Determining the number of samples in the minority class. Randomly selecting instances from the minority class and duplicating them to increase the overall count of minority class samples. Continuing the duplication process until the number of minority class samples matches that of the majority class. While ROS can help address class imbalance, it can also lead to overfitting, especially when the same observations are replicated through multiple iterations in the dataset. Studies that employ ROS have added some other unique techniques to enhance model performance. In one study, the issue of imbalanced data in binary text classification is tackled through the introduction of distributional random oversampling (Moreo et al., 2016 ). This method utilizes the distributional hypothesis, which posits that the meaning of a feature is shaped by its distribution across extensive data corpora, in order to create synthetic minority-class documents. The results suggest that the distributed random oversampling methods enhance the accuracy of classification algorithms by creating balanced datasets. Others have used ROS to address imbalances in multilabel datasets with very good results across different classification measures (Charte et al., 2015 ). In another variation of ROS, Zhao et al. ( 2016 ) address the class imbalance through stratified random oversampling. The study found that the proposed stratified oversampling method effectively addresses the challenge of imbalanced data by generating balanced and diverse training datasets (Zhao et al., 2016 ). Others have introduced the Random OverSampling approach to balancing the minority and majority classes by creating synthetic samples by randomly walking from the real data (Zhang & Li, 2014 ). The study found that random walk oversampling statistically performs much better than alternative methods on imbalanced datasets when implementing common baseline algorithms (p. 99). 3.2. SMOTE SMOTE is an oversampling technique that creates synthetic samples for the minority class by interpolating between existing minority samples. SMOTE enables researchers to use synthetic elements to rebalance under-sampled data and is one of the most effective techniques to address imbalanced datasets (Almhaithawi et al., 2020 ; Branco et al., 2017 ; Chawla et al., 2002 ). As shown in Fig. 3 , SMOTE used the k-nearest neighbor closest to the data points to create synthetic samples from the 2000 fraud instances (Branco et al., 2017 , p. 18). Instead of randomly oversampling the data with replacements, SMOTE takes "each minority class sample and introduces synthetic examples… joining any/all of the k minority class nearest neighbors" (Chawla et al., 2002 , p. 327). Depending on the number of instances needed to balance the data, SMOTE will randomly choose to generate synthetic data points from the minority sample by connecting them with their nearest neighbors (Srinilta & Kanharattanachai, 2021 ; B. Sun & Chen, 2021 ). Since its publication in 2002, SMOTE has proven successful in a variety of applications across several different domains (Fernandez et al., 2018 ). SMOTE has inspired several approaches to address class imbalance and has significantly contributed to new supervised learning paradigms, including multilabel classification, incremental learning, semi-supervised learning, and multi-instance learning, among others (p. 863). SMOTE has been effectively used to address class imbalance using machine learning algorithms in the financial crime domain with excellent results (M. E. Lokanan, 2023 ; M. E. Lokanan & Sharma, 2022 ). Others have used SMOTE to address class imbalances within high-dimensional datasets (Maldonado et al., 2019 ; S. Tiwari et al., 2018 ). Maldonado et al. ( 2019 ) propose a modified version of SMOTE designed for high-dimensional binary scenarios, such as natural language processing. This change involves using a new distance metric that focuses solely on the most important features to generate synthetic observations. Similarly, Tiwari et al. ( 2015 ) apply SMOTE to investigate the effect of various resampling ratios on observed peptides and absent peptides in protein mass spectrometry data. Both studies found that class balance greatly improves the performance of machine learning models (Maldonado et al., 2019 ; A. K. Tiwari et al., 2015 ). Others studied how close the distribution of the patterns and number of neighbors generated by SMOTE is to the original distribution and found that it performed better on large rather than small datasets (Elreedy & Atiya, 2019 ). 3.3. SVMSMOTE Support Vector Machine The Synthetic Minority OverSampling Technique (SVMSMOTE) is a type of oversampling method made for datasets that are not balanced. Its goal is to fix the overfitting problems that come with regular SMOTE (AlJame et al., 2021 , pp. 4–5). The SMOTE algorithm and Support Vector Machines (SVM) are combined in SVMSMOTE to make fake samples that are less similar to the original minority class instances (Krayem et al., 2021 ; Nguyen et al., 2011 ). To use this method, the original minority class samples are used to train an SVM classifier and find the decision boundary. This is then used to find "safe" and "borderline" areas within the minority class distribution. Synthetic samples are then subsequently generated by interpolating between instances identified as "safe," thus reducing the risk of overfitting by ensuring the synthetic samples are less similar to the original minority class instances. SVMSMOTE has been widely used across various domains to address class imbalances with enhanced performance. Studies have utilized SVMSMOTE to address class imbalances in datasets pertaining to medicine, education, and cancer research. Researchers have used SVMSMOTE to conduct diagnostic tests and predict prostate cancer using multiparametric data with outstanding results (Barlow et al., 2019; Bertelli et al., 2022 ). Sujitha and Paramasivan ( 2023 ) employed SVMSMOTE to predict stages in lung diseases with enhanced performance (Sujitha & Paramasivan, 2023 ). Others have used SVMSMOTE to guess how well students will do in multiple classes using a dataset for education that has better classification performance (Ghorbani & Ghousi, 2020 ; Tariq et al., 2023 ). All of these studies show that SVMSMOTE can be used to make models work better and fix uneven data distributions in datasets from various fields. 3.4 SMOTE Tomek SMOTETomek is a fusion of SMOTE and Tomek links, aimed at balancing imbalanced datasets while reducing noise. Applying the SMOTE algorithm to oversample the minority class is the first step in the process and removing Tomek links refines the dataset (Z. Wang et al., 2019 ). This process aims to eliminate overlapping and noisy observations, thereby enhancing the quality of the dataset and the performance of the machine learning models. By combining the strengths of SMOTE in generating synthetic samples and Tomek links in cleaning the dataset, SMOTETomek creates a more balanced and refined dataset for machine learning tasks. SMOTETomek has been widely utilized in various domains to address imbalanced datasets and enhance model performance. In healthcare, SMOTETomek has been applied to build models for disease prediction, such as diabetes and hypertension, with improved accuracy and sensitivity (Fitriyani et al., 2019 ). Cancer research studies have also used SMOTETomek to address skewed data and accurately predict high-risk prostate and cervical cancer (Boratto et al., 2022 ; Ijaz et al., 2020 ; Lin et al., 2023 ; Tanimu et al., 2022 ). In computer science, SMOTETomek has been used for recommender systems and predicting software bugs, showing that it can be used in a variety of situations and is good at fixing class imbalances (Arif et al., 2024 ; Boratto et al., 2022 ). SMOTETomek has also been used effectively to address severe sample distribution imbalances in personality recognition datasets (Z. Wang et al., 2019 ). These studies collectively underscore the utility of SMOTETomek in various domains to address class imbalance issues across various datasets. 3.5 K-MeansSMOTE K-MeansSMOTE is an oversampling technique designed to address class imbalance in datasets by generating synthetic samples for the minority class using k-means clustering. This method integrates the k-means clustering algorithm with SMOTE, executed in three distinct steps: clustering, filtering, and oversampling (Chen & Zhang, 2021 ; Douzas et al., 2018 ). Initially, the algorithm employs k-means clustering to partition the minority class samples into clusters. Subsequently, synthetic samples are created for each cluster by interpolating feature values from the minority class samples within the cluster. These synthetic instances are then added to the original dataset, resulting in a rebalanced dataset (De & Prabu, 2022 ). K-MeansSMOTE has proven to be effective in oversampling methods and has improved model performance across various classification datasets (Chen & Zhang, 2021 ; Xu et al., 2021 ). In medical science, K-MeansSMOTE has been used to effectively balance data. In a recent study, the authors applied K-MeansSMOTE to eight UCI medical datasets with excellent classification scores (Xu et al., 2021 ). In other studies, such as predicting credit default, K-MeansSMOTE has been proven to be very efficient in addressing class imbalance with enhanced performance (Alam et al., 2020 ; Chen & Zhang, 2021 ; Srinivasan et al., 2024 ; T. Sun & Vasarhelyi, 2018 ). K-MeansSMOTE has also been used to handle imbalances in classifying financial distress companies and shows improved performance across various classification metrics (Aljawazneh et al., 2021 ). Another domain where K-MeansSMOTE has been effectively used to address class imbalance and enhance performance is churn prediction (De & Prabu, 2022 ). 3.6 SMOTE + ENN SMOTe + ENN is an oversampling technique devised to rectify imbalanced datasets by combining SMOTE with the edited nearest neighbor (ENN) method. The SMOTE + ENN technique operates in two sequential steps. Initially, SMOTE is applied to produce synthetic samples for the minority class. SMOTE accomplishes this by selecting a minority class sample along with its k-nearest neighbors (k-NN), then creating new synthetic samples by interpolating the feature values between the chosen sample and its neighbors. Subsequently, the ENN method is utilized to eliminate any majority-class samples that are misclassified by a k-NN classifier. ENN detects the majority of class samples misclassified by the k-NN classifier and removes them from the dataset. By amalgamating these two steps, SMOTE + ENN can simultaneously generate new synthetic samples for the minority class and eliminate misclassified majority class samples (M. E. Lokanan, 2023 ; Sisodia et al., 2017 ; F. Yang et al., 2022 ). SMOTE + ENN has proven to be effective in balancing the class distribution and enhancing classification performance. Studies have found that SMOTE + ENN consistently yields superior outcomes compared to most oversampling methods (Batista et al., 2004 ; Singh et al., 2022 ; Sisodia et al., 2017 ). SMOTE + ENN has been extensively used to address class imbalances in medical data classification (Lamari et al., 2021 ). In medicine, SMOTE + ENN has been heavily utilized in healthcare for early detection tasks, such as predicting septic shock onset and diagnosing missed abortion, with enhanced diagnostic accuracy (Xu et al., 2020 ; F. Yang et al., 2022 ). Others have used SMOTE + ENN in predicting Parkinson's disease and chronic heart failure with very good classification results (Keller & Pandey, 2021 ; K. Wang et al., 2021 ). In the financial crimes’ arena, SMOTE + ENN has been effectively used to predict fraud with enhanced performance (M. E. Lokanan, 2023 ; Mienye & Sun, 2023 ). These studies found that SMOTE + ENN was excellent at balancing the datasets and enhancing the robustness of predictive models. 3.7 ADASYN ADASYN, short for Adaptive Synthetic Sampling, is an oversampling technique specifically designed to address class imbalances in datasets. The algorithm works by generating synthetic samples for the minority class by adaptively adjusting the density of synthetic samples based on the difficulty of the classification problem (Fernandez et al., 2018 , p. 870). The basic idea behind ADASYN is to generate more synthetic samples in regions where the decision boundary of the minority class is more complex to increase the diversity of the minority class samples (Haibo He et al., 2008 ). This adaptiveness makes ADASYN particularly effective in scenarios where the imbalance between classes is substantial and the classification problem is challenging. ADASYN oversampling has been widely applied across various domains, showcasing its effectiveness in addressing class imbalances and improving model performance. In the field of medicine, researchers have used ADASYN to up-sample features to mitigate missing value concerns in cervical cancer detection and breast cancer diagnosis with exceptional accuracy (Khan et al., 2021 ; Kurniawati et al., 2018 ; Munshi, 2024 ). Others have employed ADASYN in fraud prediction. In one study, the authors found that ADASYN proved to be more advantageous than the traditional SMOTE algorithm in telecom fraud identification (Lu et al., 2020 ). In other studies, ADASYN was used to predict insurance and credit card fraud with enhanced effectiveness on a balanced data set over an unbalanced one (Cinaroglu, 2020 ; Muranda et al., 2020 ; Singh et al., 2022 ; Subudhi & Panigrahi, 2018 ). ADASYN has also proven to be effective in customer churn prediction (Rao et al., 2024 ). These applications underscore the versatility and efficiency of using ADASYN to up-sample imbalanced datasets in different domains. The foregoing review indicates that various oversampling techniques are employed to address the challenge of imbalanced datasets in machine learning. Depending on the dataset and domain, different oversampling methods may be employed. RUS aims to balance the class distribution by manipulating minority class observations across different applications with enhanced performance. The SMOTE-based methods used SMOTE as a foundation to generate synthetic samples for the minority class and combined it with additional methods to balance the classes and enhance performance. ADASYN stands out as a unique category by adjusting synthetic sample density based on the complexity of the decision boundary. In doing so, ADASYN addresses class imbalance by focusing on domains where classification appears to be more challenging. Each oversampling technique reviewed provides distinct advantages and can be chosen according to the specific attributes and characteristics of the dataset and the classification problem being addressed. 4 Addressing the Impact of Under and OverSampling 4.1 Negative and Positive Impact of Undersampling While resampling is a useful technique to address class imbalances, there are some negative impacts associated with both undersampling and oversampling. One of the problems associated with undersampling is that it discards potentially valuable information from the dataset (Branco et al., 2017 , p. 16), an approach that tends to underfit the data (Oladunni et al., 2021 , p. 3). Think of a situation where the ratio of class imbalance is 20,000:100, 40,000:100, or 50,000:100. Removing instances to match the minority and majority classes will lead to a significant loss of data. There is no way to preserve the rich information that will be randomly removed from the majority class with undersampling (Jo & Japkowicz, 2004 ). The loss of data can make the decision boundary between the minority and majority classes harder to learn from and increase the loss function (Branco et al., 2017 ; Moreo et al., 2016 ). Another important point is that the sample chosen from the majority class may be systematically biased and not representative of the population (Chakraborty et al., 2021 ). The loss of information results in poor and inaccurate classification performance on the test set. Despite these drawbacks, undersampling is a useful strategy to address class imbalance problems. The main advantage of undersampling is that it is a valuable approach to reducing the risk of the model being trained and the analysis conducted by the classifier only on the majority class (Yap et al., 2014 ). Machine learning modeling performed only on the majority class will lead to distortion in the performance accuracy (i.e., the model is only reading the majority class) (Jo & Japkowicz, 2004 ; Rao et al., 2024 ). The minority events are harder for machine learning algorithms to predict because there are only a few instances. There is a high probability, for example, that a dataset characterized by a non-fraud to fraud ratio of 10,000:100 will lead to a model trained on the majority class (non-fraud) because there are fewer instances for the algorithm to learn from the data. Undersampling tackles this issue by reducing the number of instances in the majority class to align with the minority class. With large datasets, the use of undersampling helps to reduce computational costs and improve run time (Fujiwara et al., 2020 , p. 3). Since undersampling shrinks the data, less training time is needed (p. 3). Consequently, businesses will require less storage space and time to obtain insights from analyzing the data (Zhu et al., 2017 ). To avoid scaling issues associated with excessive data, it is best to discard some to save on computational time and resource usage. 4.2 Negative and Positive Impact of Oversampling Since minority class instances are randomly selected to produce exact copies of the majority class, random oversampling increases the likelihood of overfitting, especially for higher oversampling rates (Branco et al., 2017 ; Chawla et al., 2002 ; Fujiwara et al., 2020 ). Overfitting occurs when the model achieves high accuracy on the training data (overfits) but is a poor predictor (or fails to perform well) on the unseen test data (Santos et al., 2018 ). The use of artificial data to construct rules for the replication of the minority class that are apparently accurate may end up generating the wrong synthetic instances and increasing the generalization error (Zhu et al., 2017 , p. 330). A more fundamental problem with oversampling is that it overgeneralizes the minority class with data from the majority class and leads to "overestimation of minority data" (Pan et al., 2020 , p. 1215). This strategy is particularly problematic in the case of "highly skewed class distributions where the minority class examples are very sparse, thus resulting in a greater chance of class mixture" (Branco et al., 2017 , p. 19). Considering that the model is evaluating the same samples multiple times, datasets that are highly skewed may result in increased computational costs (Zhu et al., 2017 , p. 333). Despite these drawbacks, oversampling can have a positive impact on the results (Chawla et al., 2002 ; Elreedy & Atiya, 2019 ). The reason for oversampling the data is to ensure that there are enough instances to reduce class bias in the performance evaluation (Naseriparsa & Kashani, 2014 ). A dataset that is trained on the majority class label will result in the classifier only reading the instances from the majority class in the test set. SMOTE-based techniques can directly address this problem because they do not simply oversample the data with duplicates; rather, SMOTE generates synthetic data that is slightly different from the original data (Branco et al., 2017 ; Chawla et al., 2002 ). The main idea of oversampling is to improve the classification performance of the minority class and reduce the information loss of the majority class (Gong & Gu, 2016 , p. 3). In this regard, oversampling effectively helps to provide insights on the minority class samples, which would have been overlooked otherwise. In designing studies with imbalanced data, the minority and majority classes have to be dealt with care to ensure that each class has an equal chance to be analyzed by the classifier. Oversampling improves this outcome by giving instances from the original minority class a higher chance of being equally distributed in the data (Branco et al., 2017 ; Pan et al., 2020 ). 4.3 Deep Adaptive Resampling Technique (DART) To build on and address the limitations of commonly used oversampling techniques in handling class imbalance, this study proposes a novel method: DART. DART differentiates itself from existing resampling strategies—such as SMOTE, ADASYN, and their hybrid variants—by integrating two key innovations: (1) the generative power of deep learning through Variational Autoencoders (VAEs) and (2) adaptive sampling via pseudo-labeling of ambiguous instances. Traditional methods like SMOTE (Chawla et al., 2002 ) and ADASYN (Haibo He et al., 2008 ) use interpolation-based strategies that assume linear relationships in the minority class feature space. While effective in generating additional samples, these techniques often produce synthetic instances that lack diversity or fail to capture the complex, nonlinear structure of real-world data (Branco et al., 2017 ; Elreedy & Atiya, 2019 ). Furthermore, they do not account for class overlap or decision boundary uncertainty—issues that frequently arise in high-dimensional, noisy domains. Hybrid approaches like SMOTE-ENN or SMOTE-Tomek attempt to filter noisy samples but still rely on static resampling heuristics, limiting their adaptability to dynamic learning conditions. DART addresses these limitations through a two-stage architecture: Deep Generative Modelling with VAEs : Unlike traditional methods that rely on local interpolation (e.g., SMOTE) or density-based sampling (e.g., ADASYN), DART leverages VAEs to learn a rich latent representation of the minority class. By capturing the underlying structure of the minority class in a lower-dimensional space, VAEs enable the generation of synthetic samples that are both realistic and diverse. These samples reflect the global characteristics of the data rather than being confined to local neighborhoods. As a result, DART produces more informative and representative examples—rather than mere duplications or linear interpolations—which enhance robustness in the presence of noise and nonlinearity. Adaptive Pseudo-Labeling : DART incorporates a pseudo-labeling strategy using a Random Forest classifier to identify ambiguous majority class instances near the decision boundary (with predicted probabilities in a confidence band, e.g., 0.45–0.55). These borderline instances are selectively reassigned to the minority class based on model confidence, thereby improving the class boundary definition, and enhancing generalization (An & Suh, 2020 ). By combining these components, DART moves beyond static, rule-based sampling approaches and introduces a dynamic, data-aware framework that responds to the structure and uncertainty within the data itself. Figure 1 illustrates this dual mechanism: synthetic samples (blue triangles) generated via VAEs expand the minority class, while pseudo-labeled samples (orange squares) refine decision boundaries, all while preserving the distributional integrity of the majority class (light gray). The existing literature on class imbalance reveals a notable paucity of studies that have integrated deep generative modeling with adaptive sampling techniques. Although previous research has investigated VAE-based data augmentation (e.g., Liu et al., 2023 ) and employed classifiers for the detection of borderline instances (e.g., Ghorbani & Ghousi, 2020 ), these methodologies are typically implemented in isolation. In contrast, DART represents a distinctive synthesis of these strategies, providing a cohesive framework that not only enhances minority representation but also improves classification accuracy in proximity to the decision boundary. Consequently, DART emerges as a novel, modular, and scalable solution that can be seamlessly incorporated into any supervised learning pipeline. DART is particularly advantageous in high-stakes domains such as finance, healthcare, and fraud detection, where imbalanced data and classification inaccuracies can have profound implications. Additionally, the adaptive architecture of DART positions it as a significant advancement in the domain of imbalanced learning. 5 Research Design 5.1 Data Collection This study draws on two datasets focused on credit and loan default prediction. The credit default dataset , sourced from the UCI Machine Learning Repository, contains detailed information on credit card clients, including their demographic characteristics, credit usage behaviour, repayment history, and default payment status. 1 The second dataset, used for loan default prediction , was obtained from Kaggle and provides a complementary perspective by capturing borrower-level financial and transactional information relevant to loan performance. Together, these datasets enable a comprehensive evaluation of resampling techniques across distinct but related domains of consumer credit risk. 5.1.2 Dataset 1: Credit Card Default The credit card dataset includes information about credit card holders, such as their demographic characteristics, credit card usage, repayment history, and default payment status. The dataset consists of the following attributes: Demographic Characteristics: Demographic attributes gender, age, level of education, and marital status. Credit Card Usage: The variable LIMIT_BAL represent the amount of credit given to each cardholder. Repayment History : The dataset provides information on the repayment status of cardholders for six consecutive months (PAY_0 to PAY_6). These variables indicate whether payments were made duly or delayed, and if delayed, the duration of the delay. Bill Statements: Amount of bill statements are in reverse order from September to April (BILL_AMT1 to BILL_AMT6). Previous Payments : Includes the amount of previous payment (PAY_AMT1 to PAY_AMT6) for each cardholder over six months. Default Payment Status : The dependent variable default payment indicates whether a cardholder defaulted on their payment in the following month (1=yes, 0=no). The credit card dataset is commonly used for predictive modelling purposes, specifically to estimate the probability of default based on demographic and repayment history characteristics. Understanding credit risk and devising methods to minimize default rates is crucial for credit card issuers. There are 23 independent variables and data on 30,000 clients in the dataset. Table 1 displays the compilation of independent variables that make up the dataset. Table 1: Variables and Measurements of Credit Default dataset Variable Description Type LIMIT_BAL Amount of given credit in NT dollars (includes individual and family/supplementary credit) Continuous SEX Gender Nominal EDUCATION Level of education (1=graduate school, 2=university, 3=high school, 4=others, 5=unknown, 6=unknown) Nominal MARRIAGE Marital status (1=married, 2=single, 3=others) Nominal AGE Age in years Continuous PAY_0 to PAY_6 Repayment status from April to September 2005 (-1=pay duly, 1=payment delay for one month, 2=payment delay for two months, … 8=payment delay for eight months, 9=payment delay for nine months and above) Ordinal BILL_AMT1 to BILL_AMT6 Amount of bill statement from April to September 2005 in NT dollars Continuous PAY_AMT1 to PAY_AMT6 Amount of previous payment from April to September 2005 in NT dollars Continuous 5.1.3 Dataset 2: Load Default The loan default dataset used in this study captures a broad range of borrower and loan-related characteristics. The dataset comprises 40 features and 855,968 observations. As shown in Table 2 , the data can be grouped into five major categories: (1) Loan and credit information, including loan amount, interest rate, term, and purpose; (2) Borrower demographics and employment details, such as income, homeownership, and employment length; (3) Loan timeline and historical activities, which tracks issuance dates and the borrower’s earliest credit line; (4) Credit history and utilization metrics, including the debt-to-income ratio, revolving balances, and account delinquencies; and (5) Loan payment, recovery, and policy features, covering payment behaviour, recoveries, late fees, and compliance with lending policies. Given the dataset’s large size (over 855,000 records), processing the full set required substantial computational resources and time. To address these issues, stratified sampling was employed to reduce the dataset to 30,000 observations while preserving the original distribution of the target variable (loan default vs. non-default). The stratified sample distribution ensured balanced representation of both majority and minority classes, making the data more manageable for training and evaluating machine learning models. Table 2 Variables and Measurements of Loan Default Dataset Variable(s) Description Type loan_amnt, funded_amnt, funded_amnt_inv Loan amount applied for, amount funded, and investor-funded amount Continuous term, int_rate, installment Loan term (months), interest rate, and monthly installment payment Continuous grade, sub_grade, purpose Loan grade, sub-grade, and purpose of loan (e.g., car, education) Categorical emp_length, home_ownership, annual_inc Employment duration, homeownership status, and annual income Mixed (Nominal/Continuous) verification_status, application_type Income verification status and application type (individual or joint) Categorical issue_d, earliest_cr_line, last_pymnt_amnt Loan issuance date, earliest credit line date, and last payment amount Date / Continuous dti, delinq_2yrs, inq_last_6mths, open_acc, pub_rec, revol_bal, revol_util, total_acc Debt-to-income ratio, credit delinquencies, inquiries, open accounts, public records, revolving balance and utilization, total accounts Mixed (Continuous/Ordinal) out_prncp, out_prncp_inv, total_pymnt, total_pymnt_inv, total_rec_prncp, total_rec_int, total_rec_late_fee Loan repayment and recovery details, including outstanding principal and total payments Continuous recoveries, collection_recovery_fee Recovery and collection fees after default Continuous pymnt_plan, initial_list_status, collections_12_mths_ex_med, policy_code Payment plan status, listing status, recent collections, and policy code Categorical default Whether the borrower defaulted on the loan (1 = yes, 0 = no) Binary (Target) 5.2 Data Cleaning and Processing 5.2.1 Credit Card Dataset The dataset required minimal cleaning and pre-processing. We removed certain categories from the "Education" variable, namely "others" and the two "unknowns," to prevent interpretation issues and redundancy during feature engineering. Additionally, we identified features with a correlation above the 0.70 threshold, indicating potential multicollinearity issues. After several iterations of collinearity checks, we removed the following features from the dataset: BILL_AMT1, BILL_AMT2, BILL_AMT3, BILL_AMT4, BILL_AMT5, BILL_AMT6, PAY_3, PAY_4, PAY_5, PAY_AMT1, PAY_AMT2, PAY_AMT3, PAY_AMT4. To scale the numeric features between 0 and 1, we applied MinMaxScaler. This scaling method was chosen because certain algorithms used in the analysis, such as neural networks and k-NN, require input features to be within this range, and the numerical features did not necessarily follow a normal distribution. 5.2.2 Loan Default Dataset The loan default dataset underwent extensive cleaning and pre-processing to ensure data quality and modelling reliability. Categorical variables with ordinal significance, such as emp_length, term, and grade, were first encoded numerically to preserve order before imputation. For nominal features like home_ownership, purpose, and verification_status, one-hot encoding was applied, with one dummy variable dropped per feature to avoid multicollinearity. Features with near-zero variance, such as policy_code and acc_now_delinq, were removed due to their limited predictive value. Highly correlated features identified via the correlation matrix and variance inflation factor analysis—such as funded_amnt, funded_amnt_inv, and total_pymnt_inv—were dropped to prevent redundancy and collinearity issues. Numeric features were scaled using StandardScaler to standardize ranges across variables, which is essential for gradient-based algorithms and neural networks. Missing values were handled based on data type and context, with ordinal features transformed before imputation to avoid misleading replacements. Finally, a new variable credit_history_length was derived from earliest_cr_line and imputed where necessary to retain temporal credit behaviour information. 5.3 Resampling Techniques Applied To address class imbalance in the loan and credit default datasets, we applied a variety of commonly used undersampling and oversampling techniques. These resampling strategies were selected based on their prevalence in the literature and their suitability for binary classification tasks with highly skewed class distributions. Table 3 summarizes the techniques applied and their core mechanisms. Table 3 Summary of Resampling Techniques Used Technique Type Description Random Undersampling (RUS) Undersampling Randomly removes majority class samples to achieve class balance Tomek Links Undersampling Identifies and removes borderline majority samples paired with minority samples NearMiss Undersampling Selects majority samples that are closest to minority class instances Condensed Nearest Neighbor (CNN) Undersampling Retains only majority samples necessary for correct classification Random Oversampling Oversampling Randomly duplicates minority class instances SMOTE Oversampling Synthesizes new minority samples by interpolating between nearest neighbors SVMSMOTE Oversampling Applies SVM to generate synthetic samples along decision boundaries SMOTE-Tomek Hybrid Combines SMOTE with Tomek Links for oversampling and noise reduction SMOTE-ENN Hybrid Combines SMOTE with Edited Nearest Neighbor for enhanced noise filtering K-MeansSMOTE Oversampling Uses clustering before SMOTE to ensure diversity in synthetic samples ADASYN Oversampling Generates synthetic samples adaptively, focusing on hard-to-learn examples DART (Proposed) Hybrid Uses a Variational Autoencoder to generate synthetic samples and pseudo-labeling to adaptively include borderline majority instances 5.3.1 Over-Sampling with DART The DART technique represents an innovative two-stage hybrid oversampling method designed to mitigate class imbalance by integrating deep generative modeling with adaptive pseudo-labeling. In contrast to traditional methods such as SMOTE and ADASYN, which depend on static interpolation or fixed neighborhood heuristics, DART employs a dynamic, data-aware approach to improve the representation of minority classes and refine the delineation of class boundaries. Stage 1: Generating Synthetic Minority Samples with Variational Autoencoders (VAEs) DART begins by training a VAE on the minority class instances \(\:{X}_{\text{min}}\in\:{R}^{n\times\:d}\) , where n is the number of samples and d is the number of features. The VAE learns a latent distribution \(\:z\sim\:\mathcal{N}\left(0,I\right)\) , from which synthetic samples are drawn. These samples are then decoded using: $$\:\stackrel{\sim}{x}=\text{Decoder}\left(z\right)$$ The decoded instances \(\:\stackrel{\sim}{{X}_{\text{syn}}}\) reflect the complex, non-linear structure of the minority class and are used to balance the training data. Unlike traditional oversampling approaches, which often create limited or redundant examples, VAEs generate diverse and structurally realistic samples, improving robustness in the presence of noise and non-linearity. Stage 2: Adaptive Pseudo-Labeling for Boundary Refinement In the second stage, a Random Forest classifier is trained on the original dataset to estimate the probability that each instance belongs to the minority class: $$\:\widehat{p}\left(x\right)=P\left(y=1∣x\right)$$ In this stage, the majority of class instances with prediction probabilities within the range \(\:0.45\le\:\widehat{p}\left(x\right)\le\:0.55\) are considered ambiguous. A targeted subset of these borderline cases is pseudo-labeled as minority class and incorporated into the resampled training set. The inclusion of such cases strengthens the model's ability to distinguish between classes in regions of high uncertainty. The integration of VAE-generated samples with selectively reclassified majority instances enables DART to address both under-representation and boundary ambiguity. Minority class coverage is expanded in a way that maintains data realism, while the decision boundary is simultaneously sharpened through adaptive relabeling. Such a mechanism is particularly valuable in domains characterized by noisy, overlapping, or sparse minority class observations. 5.4 Algorithm Selection Table 4 presents the classification algorithms used to evaluate the performance of various under- and oversampling techniques. These algorithms were deliberately selected to represent a broad spectrum of machine learning categories, including ensemble methods, deep learning, linear models, and probabilistic approaches. Each classifier has a distinct mathematical foundation and learning mechanism, allowing us to examine how different algorithm types respond to class imbalance and resampling strategies. Rather than relying solely on prior literature to guide algorithm selection, we curated a balanced mix to ensure comprehensive assessment of the datasets. Using different classifications algorithms enables a robust comparison of resampling methods across diverse modelling paradigms and works to enhance the reliability and generalizability of our findings. Table 4 Algorithms Employed Category Algorithm Description Representative Formula Ensemble Learning Random Forest Combines multiple decision trees and uses majority voting for prediction. \(\:\widehat{y}=\text{mode}\left({h}_{1}\left(x\right),{h}_{2}\left(x\right),\dots\:,{h}_{T}\left(x\right)\right)\) CatBoost Gradient boosting on decision trees with categorical feature handling. Iterative updates: \(\:{F}_{t+1}\left(x\right)={F}_{t}\left(x\right)+{\eta\:}\cdot\:{h}_{t}\left(x\right)\) Deep Learning Artificial Neural Network (ANN) Uses multiple layers of neurons to learn complex patterns in data. \(\:\widehat{y}={\sigma\:}\left({W}_{2}\cdot\:\text{ReLU}\left({W}_{1}x+{b}_{1}\right)+{b}_{2}\right)\) Linear Models Logistic Regression Models log-odds of the target as a linear combination of features. \(\:\widehat{p}=\frac{1}{1+{e}^{-\left({{\beta\:}}_{0}+{{\beta\:}}_{1}{x}_{1}+\dots\:+{{\beta\:}}_{n}{x}_{n}\right)}}\) Stochastic Gradient Descent (SGD) Minimizes loss function iteratively using gradient updates. \(\:{\theta\:}={\theta\:}-{\eta\:}\cdot\:{\nabla\:}_{{\theta\:}}J\left({\theta\:}\right)\) Probabilistic Model Naive Bayes Applies Bayes' theorem assuming feature independence. \(\:P\left(y∣x\right)\propto\:P\left(y\right){\prod\:}_{i=1}^{n}P\left({x}_{i}∣y\right)\) 5.5 Evaluation Metrics We used traditional machine learning classification metrics to evaluate the performance of the models. These include accuracy, precision, recall, F1-score, and area under the ROC curve (AUC)—all of which are widely recognized and commonly used in the evaluation of classification tasks. In addition, we included the Matthews Correlation Coefficient (MCC), a more balanced metric particularly suitable for evaluating model performance on imbalanced datasets, as it considers all four components of the confusion matrix - true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). While the details of these metrics are well established in the machine learning literature and not discussed extensively here, the formula for each is presented in Table 5 for reference. Table 5 Performance Metrics Metric Description Formula Accuracy Proportion of correct predictions TP + TN/TP + TN + FP = FN Precision Proportion of true positive predictions TP/TP + FP Recall (Sensitivity) Proportion of actual positives correctly identified TP/TP + FN F1 Score Harmonic mean of precision and recall 2*Precision*Recall/Precision + Recall AUC Area under the ROC Curve ∑ i = 1 n − 12( xi + 1− xi )⋅( yi + yi + 1)/2 Matthews Correlation Coefficient (MCC) Balanced metric for binary classification \(\:\frac{TP\cdot\:TN-FP\cdot\:FN}{\sqrt{\left(TP+FP\right)\left(TP+FN\right)\left(TN+FP\right)\left(TN+FN\right)}}\) 6 Findings and Analysis Our objective in looking at these findings is to find out how the under- and oversampling techniques affect the performance of the machine learning algorithms used for data analysis. It is important to note that we are not advocating for a specific method over others. Rather, we believe that the choice of under- or oversampling techniques should be based on the type of data and the desired balance between precision, recall, and overall accuracy, depending on the specific research objectives. Our intention is to present the under- and oversampling techniques that yield the best results in addressing class imbalances using the credit default dataset. 6.1 Results of Undersampling Techniques 6.1.1 Credit Default Dataset The results in Table 7 provide a comprehensive comparison of oversampling techniques applied to credit default prediction using six machine learning algorithms. Among these methods, K-MeansSMOTE clearly emerged as the most robust and effective, delivering consistently high performance across all classifiers. Under K-MeansSMOTE, classifiers such as ANN, Gradient Descent, Logistic Regression, and CatBoost achieved accuracy scores of 0.88, precision scores above 0.90, and AUC values as high as 0.91 (ANN), with F1-scores all above 0.83. Even Naïve Bayes, which typically underperforms in imbalanced contexts, yielded a strong recall of 0.87 with K-MeansSMOTE, demonstrating its enhanced sensitivity to the minority class. These results suggest that K-MeansSMOTE not only balances the dataset effectively but also generates high-quality synthetic samples that preserve class boundaries, leading to better model generalization and discrimination (Chen & Zhang, 2021 ; De & Prabu, 2022 ; Douzas et al., 2018 ). Other methods showed moderate to inconsistent results. SMOTEENN, for example, performed reasonably well with ANN, producing F1-scores of up to 0.79 and AUC values around 0.87, but struggled with models like CatBoost and Random Forest. ROS and SMOTE achieved mid-range performance across metrics, while ADASYN and SVMSMOTE yielded the weakest outcomes, particularly in precision and AUC, indicating difficulties in generating informative minority samples. Notably, ANN consistently performed well across all resampling methods, further validating its flexibility and adaptability to imbalanced datasets (Chawla et al., 2002 ; Kurani et al., 2023 ). Table 7 Oversampling Sampling Results for Credit Default Dataset Algorithm Accuracy Precision Recall F1-Score AUC Method CatBoost 0.88 0.91 0.84 0.87 0.88 KMeansSMOTE Random Forest 0.87 0.88 0.85 0.86 0.87 KMeansSMOTE Naïve Bayes 0.81 0.78 0.87 0.82 0.81 KMeansSMOTE Logistic Regression 0.88 0.91 0.84 0.87 0.88 KMeansSMOTE Gradient Descent 0.88 0.90 0.84 0.87 0.88 KMeansSMOTE ANN 0.88 0.92 0.83 0.87 0.91 KMeansSMOTE CatBoost 0.78 0.93 0.65 0.77 0.80 SMOTEENN Random Forest 0.70 0.70 0.81 0.75 0.68 SMOTEENN Naïve Bayes 0.74 0.75 0.80 0.77 0.73 SMOTEENN Logistic Regression 0.73 0.76 0.77 0.76 0.73 SMOTEENN Gradient Descent 0.73 0.75 0.77 0.76 0.72 SMOTEENN ANN 0.78 0.84 0.74 0.79 0.87 SMOTEENN CatBoost 0.69 0.78 0.52 0.62 0.69 ROS Random Forest 0.69 0.78 0.52 0.62 0.69 ROS Naïve Bayes 0.65 0.64 0.69 0.66 0.65 ROS Logistic Regression 0.66 0.67 0.63 0.65 0.66 ROS Gradient Descent 0.66 0.66 0.64 0.65 0.66 ROS ANN 0.70 0.77 0.58 0.66 0.76 ROS CatBoost 0.70 0.79 0.54 0.64 0.70 SVMSMOTE Random Forest 0.55 0.53 0.84 0.65 0.55 SVMSMOTE Naïve Bayes 0.65 0.62 0.75 0.68 0.65 SVMSMOTE Logistic Regression 0.68 0.76 0.54 0.63 0.68 SVMSMOTE Gradient Descent 0.69 0.76 0.55 0.64 0.69 SVMSMOTE ANN 0.71 0.79 0.56 0.66 0.78 SVMSMOTE CatBoost 0.68 0.80 0.52 0.63 0.69 SMOTETomek Random Forest 0.60 0.58 0.80 0.67 0.59 SMOTETomek Naïve Bayes 0.65 0.63 0.74 0.68 0.65 SMOTETomek Logistic Regression 0.66 0.67 0.68 0.67 0.66 SMOTETomek Gradient Descent 0.64 0.65 0.65 0.65 0.64 SMOTETomek ANN 0.69 0.73 0.62 0.67 0.76 SMOTETomek CatBoost 0.68 0.78 0.51 0.62 0.68 SMOTE Random Forest 0.61 0.58 0.77 0.66 0.61 SMOTE Naïve Bayes 0.65 0.64 0.71 0.67 0.65 SMOTE Logistic Regression 0.66 0.67 0.64 0.65 0.66 SMOTE Gradient Descent 0.65 0.66 0.63 0.64 0.65 SMOTE ANN 0.68 0.80 0.48 0.60 0.76 SMOTE CatBoost 0.64 0.74 0.43 0.54 0.64 ADASYN Random Forest 0.54 0.52 0.79 0.63 0.54 ADASYN Naïve Bayes 0.61 0.59 0.70 0.64 0.61 ADASYN Logistic Regression 0.62 0.61 0.63 0.62 0.62 ADASYN Gradient Descent 0.62 0.61 0.62 0.61 0.62 ADASYN ANN 0.65 0.73 0.46 0.56 0.70 ADASYN 6.1.2 Loan Default Dataset Table 8 displays the oversampling results for the loan default dataset. The evaluation results highlight several important trends across resampling methods and machine learning classifiers. Like the results from the credit default dataset, K-MeansSMOTE consistently outperformed all other resampling techniques, producing the highest F1-Scores (up to 0.71), strong AUC values (up to 0.92), and the highest MCC of 0.73 with both CatBoost and Random Forest classifiers. Notably, CatBoost emerged as the most effective algorithm overall, maintaining top performance across nearly all resampling techniques. For example, under K-MeansSMOTE, SMOTE, and ADASYN, CatBoost achieved the best combination of accuracy, precision, AUC, and MCC. ANN and Gradient Descent classifiers, while generally performing well, showed greater variability depending on the oversampling method used—highlighting their sensitivity to data distributions. Although SMOTEENN, ROS, and SVMSMOTE provided moderate improvements in minority class recall, their precision and MCC were generally lower than those achieved with K-MeansSMOTE. ADASYN, despite performing decently in recall, struggled to balance precision and F1-Scores across most classifiers. These findings reaffirm that K-MeansSMOTE, when paired with CatBoost or Random Forest, offers the most robust and generalizable solution for handling class imbalance in loan default prediction (Aljawazneh et al., 2021 ; De & Prabu, 2022 ; Srinivasan et al., 2024 ). Table 8 Oversampling Sampling Results for Loan Default Dataset CatBoost 0.98 0.97 0.56 0.71 0.92 0.73 KMeansSMOTE Random Forest 0.97 0.98 0.55 0.70 0.90 0.72 KMeansSMOTE Naïve Bayes 0.92 0.35 0.65 0.46 0.85 0.44 KMeansSMOTE Logistic Regression 0.87 0.25 0.70 0.37 0.87 0.37 KMeansSMOTE Gradient Descent 0.86 0.24 0.70 0.36 0.87 0.36 KMeansSMOTE ANN 0.95 0.52 0.57 0.54 0.86 0.52 KMeansSMOTE CatBoost 0.97 0.81 0.59 0.68 0.92 0.68 SMOTEENN Random Forest 0.97 0.82 0.57 0.67 0.91 0.67 SMOTEENN Naïve Bayes 0.95 0.56 0.61 0.58 0.88 0.56 SMOTEENN Gradient Descent 0.81 0.20 0.79 0.32 0.90 0.33 SMOTEENN Logistic Regression 0.80 0.19 0.82 0.31 0.90 0.33 SMOTEENN ANN 0.92 0.36 0.63 0.46 0.87 0.44 SMOTEENN CatBoost 0.97 0.78 0.59 0.67 0.92 0.66 ROS Random Forest 0.93 0.42 0.65 0.51 0.90 0.49 ROS Naïve Bayes 0.95 0.57 0.60 0.59 0.89 0.56 ROS Gradient Descent 0.86 0.24 0.76 0.37 0.90 0.38 ROS Logistic Regression 0.87 0.25 0.76 0.38 0.91 0.39 ROS ANN 0.95 0.56 0.58 0.57 0.86 0.54 ROS CatBoost 0.97 0.96 0.56 0.70 0.92 0.72 SVMSMOTE Random Forest 0.97 0.96 0.56 0.71 0.91 0.72 SVMSMOTE Naïve Bayes 0.95 0.50 0.60 0.54 0.87 0.52 SVMSMOTE Gradient Descent 0.87 0.27 0.74 0.39 0.89 0.39 SVMSMOTE Logistic Regression 0.90 0.32 0.70 0.44 0.90 0.43 SVMSMOTE ANN 0.95 0.54 0.57 0.55 0.85 0.53 SVMSMOTE CatBoost 0.98 0.97 0.55 0.71 0.92 0.72 SMOTETomek Random Forest 0.97 0.97 0.55 0.70 0.90 0.72 SMOTETomek Naïve Bayes 0.96 0.60 0.60 0.60 0.88 0.57 SMOTETomek Gradient Descent 0.89 0.28 0.74 0.41 0.90 0.41 SMOTETomek Logistic Regression 0.87 0.26 0.75 0.39 0.90 0.39 SMOTETomek ANN 0.94 0.47 0.58 0.52 0.86 0.50 SMOTETomek CatBoost 0.98 0.98 0.56 0.71 0.92 0.73 SMOTE Random Forest 0.98 0.98 0.55 0.71 0.90 0.73 SMOTE Naïve Bayes 0.97 0.60 0.60 0.60 0.88 0.57 SMOTE Gradient Descent 0.86 0.24 0.76 0.37 0.90 0.38 SMOTE Logistic Regression 0.87 0.26 0.75 0.39 0.90 0.39 SMOTE ANN 0.94 0.48 0.60 0.53 0.86 0.50 SMOTE CatBoost 0.98 0.97 0.56 0.71 0.92 0.72 ADASYN Random Forest 0.97 0.95 0.56 0.70 0.90 0.72 ADASYN Naïve Bayes 0.95 0.55 0.61 0.58 0.88 0.55 ADASYN Gradient Descent 0.82 0.2 0.80 0.32 0.90 0.34 ADASYN Logistic Regression 0.81 0.20 0.81 0.32 0.90 0.34 ADASYN ANN 0.94 0.46 0.60 0.52 0.87 0.49 ADASYN For both the credit and loan default datasets, K-MeansSMOTE emerged as the most effective oversampling technique, consistently delivering superior performance across a wide range of evaluation metrics, including F1-Score, AUC, and MCC. Among the classifiers tested, CatBoost stood out as the top-performing algorithm on both datasets, demonstrating robust precision, strong recall, and excellent generalization capability across different resampling strategies (J. T. Hancock & Khoshgoftaar, 2020 ; Huang et al., 2019 ). The combination of K-MeansSMOTE with CatBoost offers a powerful and scalable approach to addressing class imbalance and achieving high predictive performance in credit risk modelling tasks. These findings suggest that practitioners dealing with imbalanced financial datasets should strongly consider the K-MeansSMOTE with CatBoost pairing for reliable and accurate classification outcomes. 6.2 Results of Undersampling Techniques 6.2.1 Credit Default Dataset Table 9 provides the results for the undersampling techniques of the credit default dataset. The results reveal that NearMiss outperformed other methods in balancing predictive performance across models. Particularly, when combined with ANN, NearMiss achieved the best overall performance, with an F1-Score of 0.81, AUC of 0.87, and strong precision and recall (0.83 and 0.80, respectively). These results suggests that NearMiss is an effective technique to preserve informative patterns from the minority class while reducing the majority class (Ha & Lee, 2016 ; Mqadi et al., 2021 ). TomekLinks, in contrast, exhibited lower recall across all classifiers—highlighting its limited utility in retaining true positive instances (Ai-jun & Peng, 2020 ; Devi et al., 2017 ). For example, while CatBoost under TomekLinks attained a high precision of 0.76, its recall was only 0.31, resulting in a low F1-score of 0.44. Random undersampling yielded moderate performance across classifiers, with F1-scores hovering around 0.62–0.66, suggesting it may be too aggressive in discarding data. CNN, similarly, underperformed in recall, especially with models like ANN and Gradient Descent, where recall dropped to 0.42 and 0.48, respectively. Table 9 Undersampling Sampling Results for Credit Default Dataset Algorithm Accuracy Precision Recall F1_Score AUC Method CatBoost 0.82 0.76 0.31 0.44 0.64 TomekLinks Random Forest 0.79 0.57 0.52 0.54 0.70 TomekLinks Naïve Bayes 0.80 0.58 0.49 0.53 0.69 TomekLinks Logistic Regression 0.81 0.74 0.32 0.45 0.64 TomekLinks Gradient Descent 0.80 0.70 0.25 0.37 0.61 TomekLinks ANN 0.82 0.74 0.34 0.47 0.77 TomekLinks CatBoost 0.76 0.78 0.72 0.75 0.76 NearMiss Random Forest 0.76 0.79 0.72 0.75 0.76 NearMiss Naïve Bayes 0.72 0.83 0.57 0.68 0.72 NearMiss Logistic Regression 0.68 0.70 0.63 0.66 0.68 NearMiss Gradient Descent 0.68 0.70 0.63 0.66 0.68 NearMiss ANN 0.82 0.83 0.80 0.81 0.87 NearMiss CatBoost 0.69 0.79 0.51 0.62 0.69 Random Random Forest 0.69 0.8 0.51 0.62 0.69 Random Naïve Bayes 0.65 0.64 0.69 0.66 0.65 Random Logistic Regression 0.66 0.67 0.62 0.64 0.66 Random Gradient Descent 0.65 0.65 0.67 0.66 0.65 Random ANN 0.70 0.77 0.56 0.65 0.76 Random CatBoost 0.66 0.72 0.43 0.54 0.64 CNN Random Forest 0.66 0.71 0.46 0.56 0.65 CNN Naïve Bayes 0.64 0.62 0.58 0.60 0.64 CNN Logistic Regression 0.66 0.69 0.46 0.55 0.64 CNN Gradient Descent 0.65 0.66 0.48 0.56 0.63 CNN ANN 0.66 0.73 0.42 0.53 0.70 CNN 6.2.2 Loan Default Dataset As shown in Table 10 , the results from applying various undersampling techniques to the loan default dataset reveal substantial variation in model performance depending on the method used. TomekLinks emerged as the most effective undersampling technique overall, delivering consistently high scores across classifiers such as CatBoost, Random Forest, Logistic Regression, and ANN. These combinations achieved F1-scores up to 0.71 and MCC values as high as 0.73, indicating balanced precision-recall trade-offs and strong alignment between predicted and true labels. Special attention is given to the F1-score and the MCC, which are more appropriate for evaluating imbalanced datasets. In contrast, NearMiss performed poorly across all classifiers, with significant drops in F1-scores (as low as 0.13 to 0.25) and MCC values barely exceeding 0.24. Random undersampling produced moderate results, with Naïve Bayes achieving the best MCC (0.63) and F1-score (0.64), while most other classifiers struggled. CNN showed stronger performance, especially when paired with Random Forest and CatBoost, achieving F1-scores of 0.71 and MCC values of 0.72 and 0.73 respectively—comparable to those attained by TomekLinks. Despite CNN’s competitive performance, the consistent superiority of TomekLinks across nearly all evaluation metrics and models establishes it as the most robust and reliable undersampling method for the loan default dataset (see Alamri & Ykhlef, 2024 ; Devi et al., 2017 ; Pereira et al., 2020 ). Table 10 Undersampling Sampling Results for Loan Default Dataset Algorithm Accuracy Precision Recall F1_Score AUC MCC Method CatBoost 0.98 0.98 0.56 0.71 0.92 0.73 TomekLinks Random Forest 0.98 1.00 0.55 0.71 0.89 0.73 TomekLinks Naïve Bayes 0.96 0.60 0.60 0.60 0.89 0.57 TomekLinks Logistic Regression 0.97 0.97 0.52 0.67 0.90 0.69 TomekLinks Gradient Descent 0.97 0.98 0.50 0.67 0.89 0.69 TomekLinks ANN 0.96 0.65 0.56 0.60 0.86 0.58 TomekLinks CatBoost 0.42 0.07 0.86 0.14 0.80 0.12 NearMiss Random Forest 0.37 0.07 0.88 0.13 0.80 0.10 NearMiss Naïve Bayes 0.77 0.15 0.68 0.25 0.73 0.24 NearMiss Logistic Regression 0.49 0.09 0.86 0.16 0.80 0.15 NearMiss Gradient Descent 0.47 0.08 0.89 0.15 0.81 0.15 NearMiss ANN 0.36 0.07 0.90 0.13 0.72 0.11 NearMiss CatBoost 0.85 0.23 0.78 0.36 0.92 0.37 Random Random Forest 0.87 0.25 0.74 0.38 0.91 0.38 Random Naïve Bayes 0.97 0.73 0.57 0.64 0.89 0.63 Random Logistic Regression 0.85 0.23 0.76 0.35 0.90 0.36 Random Gradient Descent 0.83 0.21 0.76 0.33 0.89 0.34 Random ANN 0.77 0.16 0.79 0.27 0.87 0.29 Random CatBoost 0.97 0.95 0.56 0.71 0.92 0.72 CNN Random Forest 0.98 0.98 0.55 0.71 0.90 0.73 CNN Naïve Bayes 0.97 0.78 0.57 0.66 0.90 0.65 CNN Logistic Regression 0.97 0.93 0.52 0.67 0.90 0.69 CNN Gradient Descent 0.97 0.77 0.53 0.63 0.85 0.62 CNN ANN 0.93 0.39 0.64 0.48 0.86 0.46 CNN 6.3 DART Oversampling Table 11 presents the results of the DART oversampling method applied to both the credit and loan default datasets. As evidenced by the results, DART performed exceptionally well across all classifiers and evaluation metrics. Given the imbalanced nature of the datasets, traditional accuracy alone is not a sufficient indicator of model performance. Therefore, emphasis is placed on more robust evaluation metrics—F1-Score, AUC, and the MCC —which are better suited for imbalanced classification problems (Y. Bao et al., 2022 ; M. Lokanan & Sharma, 2024 ). Remarkably, DART achieved scores exceeding 0.98 for all three of these metrics across all classifiers in both datasets. The consistency of these results underscores DART’s effectiveness in generating high-quality synthetic samples that support generalization across multiple algorithms. Notably, CatBoost and Random Forest emerged as the top-performing classifiers under the DART algorithm, reinforcing their ability to capitalize on the enhanced balance and diversity introduced by the resampling technique. These findings highlight DART as a state-of-the-art oversampling solution with superior performance and broad applicability in predictive modelling tasks involving class imbalance. Table 11 DART Oversampling Results DART Credit Default Algorithm Accuracy Precision Recall F1-Score AUC MCC Method CatBoost 0.99 0.100 0.99 0.99 1.00 0.99 DART Random Forest 0.99 0.99 0.99 0.99 1.00 0.98 DART Naïve Bayes 0.98 1.00 0.96 0.98 0.99 0.96 DART Logistic Regression 0.99 0.99 0.99 0.99 1.00 0.97 DART Gradient Descent 0.99 0.99 0.99 0.99 1.00 0.98 DART ANN 0.98 0.99 0.97 0.98 1.00 0.97 DART DART Loan Algorithm Accuracy Precision Recall F1-Score AUC MCC Method CatBoost 0.99 0.998 0.99 0.99 1.00 0.99 DART Random Forest 0.99 0.998 0.99 0.99 1.00 0.99 DART Naïve Bayes 0.98 1.00 0.95 0.98 0.99 0.96 DART Logistic Regression 0.98 0.98 0.99 0.98 1.00 0.97 DART Gradient Descent 0.99 0.98 0.99 0.99 1.00 0.97 DART ANN 0.99 0.99 0.99 0.99 1.00 0.98 DART 7 Discussion and Conclusion The aim of this study was to address the pervasive issue of class imbalance in predictive modeling by evaluating the effectiveness of both advanced undersampling and oversampling techniques. Using two large and highly imbalanced datasets—one on credit default and the other on loan default—the study systematically assessed how different resampling strategies impact model performance across a variety of machine learning algorithms, including ensemble methods, probabilistic classifiers, linear models, and deep learning approaches. A central component of this investigation was the introduction of a novel oversampling technique, DART, which was benchmarked against established methods to determine its robustness, adaptability, and generalizability across different datasets and model types. In the case of undersampling, NearMiss emerged as the most effective technique for the credit default dataset, delivering balanced F1-scores and AUC values across various classifiers. These result corroborates prior research emphasizing NearMiss’s utility in enhancing classifier performance through selective removal of overlapping majority samples (Mqadi et al., 2021 ; Alamsyah et al., 2022 ). Meanwhile, TomekLinks outperformed all other undersampling methods in the loan default dataset, showing superior results in terms of F1-score and MCC, especially when paired with CatBoost and Random Forest (Ai-jun & Peng, 2020 ; Devi et al., 2017 ; Pereira et al., 2020 ). These findings are consistent with the literature highlighting TomekLinks’ ability to clean noisy boundary points and clarify decision margins (Devi et al., 2017 ; Zeng et al., 2016 ). When analyzing oversampling techniques, K-MeansSMOTE consistently delivered the best results across all classifiers in both datasets, based on a range of established evaluation metrics. Oversampling sampling with K-MeansSMOTE achieved high scores in F1, AUC, and MCC—metrics that are particularly suited for evaluating imbalanced datasets (Douzas et al., 2018 ; Chen & Zhang, 2021 ). These findings support the use of K-MeansSMOTE in domains like finance and healthcare, where class boundaries are complex and misclassification is costly (Alam et al., 2020 ; Douzas et al., 2018 ; Srinivasan et al., 2024 ). SMOTE + ENN also performed well in certain contexts, particularly by enhancing precision and reducing class overlap, corroborating previous findings in medical and fraud detection research (M. E. Lokanan, 2023 ; Sisodia et al., 2017 ; F. Yang et al., 2022 ). Conversely, ADASYN tended to underperform in both datasets, reinforcing concerns from prior studies about its sensitivity to noise in highly imbalanced scenarios (Haibo He et al., 2008 ; Singh et al., 2022 ). Most notably, DART outperformed all traditional oversampling methods across all evaluation metrics. With F1-scores, AUC, and MCC values consistently above 0.97 for both datasets, DART demonstrated an exceptional ability to synthesize informative samples while preserving the underlying distribution of the data. These results underscore the advantage of deep adaptive learning techniques in resampling tasks and open new avenues for improving classification in highly imbalanced domains. Overall, the results reaffirm the critical role of choosing appropriate resampling techniques based on dataset characteristics, domain constraints, and the performance metric most aligned with the application context (Branco et al., 2017 ; Pan et al., 2020 ). Our study contributes to the growing body of knowledge by empirically validating and benchmarking resampling techniques across diverse classifiers and datasets. Moreover, our study introduces a novel, highly effective oversampling techniques —DART—which when tested across two different datasets offers a promising new direction for future work in imbalanced classification. 7.1 Limitations and Future Research Despite its contributions, this study is not without limitations. First, while DART demonstrates promising results, its current implementation is limited to oversampling the minority class. As such, it does not address scenarios where undersampling or hybrid sampling may be more appropriate, particularly in cases where the majority class contains significant noise or redundancy. Limiting DART to oversampling of the minority class limits it applicability to tasks where generating synthetic minority examples is suitable and may not perform optimally in datasets that would benefit from a combination of oversampling and majority class reduction. Future work should explore extending DART to support undersampling strategies or integrating it into more flexible hybrid frameworks that adapt to varying data distributions. Second, while this paper employed a range of widely accepted evaluation metrics—including accuracy, precision, recall, F1-score, AUC, and the MCC—it did not exhaustively explore the full range of experimental setups and performance measures commonly used in predictive machine learning research. We employed a standard train/test split approach to evaluate model performance; however, alternative validation techniques such as k-fold cross-validation and bootstrapped sample could offer more robust estimates and help mitigate variance due to data partitioning. Future research would benefit from incorporating cross-validation, bootstrapped sampling, or other advanced resampling techniques to enhance the reliability and generalizability of model performance. Declarations Conflict of interest There is no conflict of interest. Funding statement There is no funding associated with this paper. Acknowledgments I would like to acknowledge the contribution made by my graduate research assistant, Vikas Maddhesia, to this project. Data Availability Data for this paper will be available upon request. References Ai-jun, L., & Peng, Z. (2020). Research on Unbalanced Data Processing Algorithm Base Tomeklinks-Smote. Proceedings of the 2020 3rd International Conference on Artificial Intelligence and Pattern Recognition , 13–17. https://doi.org/10.1145/3430199.3430222 Alam, T. M., Shaukat, K., Hameed, I. A., Luo, S., Sarwar, M. U., Shabbir, S., Li, J., & Khushi, M. (2020). An Investigation of Credit Card Default Prediction in the Imbalanced Datasets. IEEE Access , 8 , 201173–201198. https://doi.org/10.1109/ACCESS.2020.3033784 Alamri, M., & Ykhlef, M. (2024). Hybrid Undersampling and Oversampling for Handling Imbalanced Credit Card Data. IEEE Access , 12 , 14050–14060. https://doi.org/10.1109/ACCESS.2024.3357091 Alamsyah, A. R. B., Anisa, S. R., Belinda, N. S., & Setiawan, A. (2022). SMOTE and Nearmiss Methods for Disease Classification with Unbalanced Data: Case Study: IFLS 5. Proceedings of The International Conference on Data Science and Official Statistics , 2021 (1), 305–314. https://doi.org/10.34123/icdsos.v2021i1.240 AlJame, M., Imtiaz, A., Ahmad, I., & Mohammed, A. (2021). Deep forest model for diagnosing COVID-19 from routine blood tests. Scientific Reports , 11 (1), 16682. https://doi.org/10.1038/s41598-021-95957-w Aljawazneh, H., Mora, A. M., Garcia-Sanchez, P., & Castillo-Valdivieso, P. A. (2021). Comparing the Performance of Deep Learning Methods to Predict Companies’ Financial Failure. IEEE Access , 9 , 97010–97038. https://doi.org/10.1109/ACCESS.2021.3093461 Almhaithawi, D., Jafar, A., & Aljnidi, M. (2020). Example-dependent cost-sensitive credit cards fraud detection using SMOTE and Bayes minimum risk. SN Applied Sciences , 2 (9), 1574. https://doi.org/10.1007/s42452-020-03375-w Alsowail, R. A. (2022). An Insider Threat Detection Model Using One-Hot Encoding and Near-Miss Under-Sampling Techniques. In M. S. Uddin, P. K. Jamwal, & J. C. Bansal (Eds.), Proceedings of International Joint Conference on Advances in Computational Intelligence (pp. 183–196). Springer Nature Singapore. https://doi.org/10.1007/978-981-19-0332-8_13 An, B., & Suh, Y. (2020). Identifying financial statement fraud with decision rules obtained from Modified Random Forest. Data Technologies and Applications , 54 (2), 235–255. https://doi.org/10.1108/DTA-11-2019-0208 Araf, I., Idri, A., & Chairi, I. (2024). Cost-sensitive learning for imbalanced medical data: A review. Artificial Intelligence Review , 57 (4), 80. https://doi.org/10.1007/s10462-023-10652-8 Arif, M., Fang, G., Fida, H., Musleh, S., Yu, D.-J., & Alam, T. (2024). iMRSAPred: Improved Prediction of Anti-MRSA Peptides Using Physicochemical and Pairwise Contact-Energy Properties of Amino Acids. ACS Omega , 9 (2), 2874–2883. https://doi.org/10.1021/acsomega.3c08303 Bagui, S., & Li, K. (2021). Resampling imbalanced data for network intrusion detection datasets. Journal of Big Data , 8 (1), 6. https://doi.org/10.1186/s40537-020-00390-x Bansal, A., & Jain, A. (2021). Analysis of Focussed Under-Sampling Techniques with Machine Learning Classifiers. 2021 IEEE/ACIS 19th International Conference on Software Engineering Research, Management and Applications (SERA) , 91–96. https://doi.org/10.1109/SERA51205.2021.9509270 Bao, L., Juan, C., Li, J., & Zhang, Y. (2016). Boosted Near-miss Under-sampling on SVM ensembles for concept detection in large-scale imbalanced datasets. Neurocomputing , 172 , 198–206. https://doi.org/10.1016/j.neucom.2014.05.096 Bao, Y., Hilary, G., & Ke, B. (2022). Artificial Intelligence and Fraud Detection. In V. Babich, J. R. Birge, & G. Hilary (Eds.), Innovative Technology at the Interface of Finance and Operations (Vol. 11, pp. 223–247). Springer International Publishing. https://doi.org/10.1007/978-3-030-75729-8_8 Barlow, Mao, & Khushi. (2019). Predicting High-Risk Prostate Cancer Using Machine Learning Methods. Data , 4 (3), 129. https://doi.org/10.3390/data4030129 Basit, M. S., Khan, A., Farooq, O., Khan, Y. U., & Shameem, M. (2022). Handling Imbalanced and Overlapped Medical Datasets: A Comparative Study. 2022 5th International Conference on Multimedia, Signal Processing and Communication Technologies (IMPACT) , 1–7. https://doi.org/10.1109/IMPACT55510.2022.10029111 Batista, G. E. A. P. A., Prati, R. C., & Monard, M. C. (2004). A study of the behavior of several methods for balancing machine learning training data. ACM SIGKDD Explorations Newsletter , 6 (1), 20–29. https://doi.org/10.1145/1007730.1007735 Bauder, R., & Khoshgoftaar, T. (2018). Medicare Fraud Detection Using Random Forest with Class Imbalanced Big Data. 2018 IEEE International Conference on Information Reuse and Integration (IRI) , 80–87. https://doi.org/10.1109/IRI.2018.00019 Bertelli, E., Mercatelli, L., Marzi, C., Pachetti, E., Baccini, M., Barucci, A., Colantonio, S., Gherardini, L., Lattavo, L., Pascali, M. A., Agostini, S., & Miele, V. (2022). Machine and Deep Learning Prediction Of Prostate Cancer Aggressiveness Using Multiparametric MRI. Frontiers in Oncology , 11 , 802964. https://doi.org/10.3389/fonc.2021.802964 Boratto, L., Carta, S., Iguider, W., Mulas, F., & Pilloni, P. (2022). Fair performance-based user recommendation in eCoaching systems. User Modeling and User-Adapted Interaction , 32 (5), 839–881. https://doi.org/10.1007/s11257-022-09339-6 Branco, P., Torgo, L., & Ribeiro, R. P. (2017). A Survey of Predictive Modeling on Imbalanced Domains. ACM Computing Surveys , 49 (2), 1–50. https://doi.org/10.1145/2907070 Chakraborty, J., Majumder, S., & Menzies, T. (2021). Bias in machine learning software: Why? how? what to do? Proceedings of the 29th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering , 429–440. https://doi.org/10.1145/3468264.3468537 Chaplot, A., Choudhary, N., & Jain, K. (2019). A Review on Data Level Approaches for Managing Imbalanced Classification Problem . 6 (2), 91–97. Charte, F., Rivera, A. J., del Jesus, M. J., & Herrera, F. (2015). Addressing imbalance in multilabel classification: Measures and random resampling algorithms. Neurocomputing , 163 , 3–16. https://doi.org/10.1016/j.neucom.2014.08.091 Chawla, N. V., Bowyer, K. W., Hall, L. O., & Kegelmeyer, W. P. (2002). SMOTE: Synthetic Minority Over-sampling Technique. Journal of Artificial Intelligence Research , 16 , 321–357. https://doi.org/10.1613/jair.953 Chen, Y., & Zhang, R. (2021). Research on Credit Card Default Prediction Based on k-Means SMOTE and BP Neural Network. Complexity , 2021 , 1–13. https://doi.org/10.1155/2021/6618841 Cinaroglu, S. (2020). Modelling unbalanced catastrophic health expenditure data by using machine‐learning methods. Intelligent Systems in Accounting, Finance and Management , 27 (4), 168–181. https://doi.org/10.1002/isaf.1483 Dal Pozzolo, A., Caelen, O., Le Borgne, Y.-A., Waterschoot, S., & Bontempi, G. (2014). Learned lessons in credit card fraud detection from a practitioner perspective. Expert Systems with Applications , 41 (10), 4915–4928. https://doi.org/10.1016/j.eswa.2014.02.026 De, S., & Prabu, P. (2022). A Sampling-Based Stack Framework for Imbalanced Learning in Churn Prediction. IEEE Access , 10 , 68017–68028. https://doi.org/10.1109/ACCESS.2022.3185227 Devi, D., Biswas, S. kr., & Purkayastha, B. (2017). Redundancy-driven modified Tomek-link based undersampling: A solution to class imbalance. Pattern Recognition Letters , 93 , 3–12. https://doi.org/10.1016/j.patrec.2016.10.006 Douzas, G., Bacao, F., & Last, F. (2018). Improving imbalanced learning through a heuristic oversampling method based on k-means and SMOTE. Information Sciences , 465 , 1–20. https://doi.org/10.1016/j.ins.2018.06.056 Elreedy, D., & Atiya, A. F. (2019). A Comprehensive Analysis of Synthetic Minority Oversampling Technique (SMOTE) for handling class imbalance. Information Sciences , 505 , 32–64. https://doi.org/10.1016/j.ins.2019.07.070 Fernandez, A., Garcia, S., Herrera, F., & Chawla, N. V. (2018). SMOTE for Learning from Imbalanced Data: Progress and Challenges, Marking the 15-year Anniversary. Journal of Artificial Intelligence Research , 61 , 863–905. https://doi.org/10.1613/jair.1.11192 Fitriyani, N. L., Syafrudin, M., Alfian, G., & Rhee, J. (2019). Development of Disease Prediction Model Based on Ensemble Learning Approach for Diabetes and Hypertension. IEEE Access , 7 , 144777–144789. https://doi.org/10.1109/ACCESS.2019.2945129 Fujiwara, K., Huang, Y., Hori, K., Nishioji, K., Kobayashi, M., Kamaguchi, M., & Kano, M. (2020). Over- and Under-sampling Approach for Extremely Imbalanced and Small Minority Data Problem in Health Record Analysis. Frontiers in Public Health , 8 , 178. https://doi.org/10.3389/fpubh.2020.00178 Ghorbani, R., & Ghousi, R. (2020). Comparing Different Resampling Methods in Predicting Students’ Performance Using Machine Learning Techniques. IEEE Access , 8 , 67899–67911. https://doi.org/10.1109/ACCESS.2020.2986809 Gong, C., & Gu, L. (2016). A Novel SMOTE-Based Classification Approach to Online Data Imbalance Problem. Mathematical Problems in Engineering , 2016 , 1–14. https://doi.org/10.1155/2016/5685970 Ha, J., & Lee, J.-S. (2016). A New Under-Sampling Method Using Genetic Algorithm for Imbalanced Data Classification. Proceedings of the 10th International Conference on Ubiquitous Information Management and Communication , 1–6. https://doi.org/10.1145/2857546.2857643 Haibo He, Yang Bai, Garcia, E. A., & Shutao Li. (2008). ADASYN: Adaptive synthetic sampling approach for imbalanced learning. 2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence) , 1322–1328. https://doi.org/10.1109/IJCNN.2008.4633969 Hancock, J., Khoshgoftaar, T. M., & Johnson, J. M. (2022). The Effects of Random Undersampling for Big Data Medicare Fraud Detection. 2022 IEEE International Conference on Service-Oriented System Engineering (SOSE) , 141–146. https://doi.org/10.1109/SOSE55356.2022.00023 Hancock, J. T., & Khoshgoftaar, T. M. (2020). CatBoost for big data: An interdisciplinary review. Journal of Big Data , 7 (1), 94. https://doi.org/10.1186/s40537-020-00369-8 Hasanin, T., Khoshgoftaar, T. M., Leevy, J. L., & Bauder, R. A. (2019). Severely imbalanced Big Data challenges: Investigating data sampling approaches. Journal of Big Data , 6 (1), 107. https://doi.org/10.1186/s40537-019-0274-4 Hernandez, J., Carrasco-Ochoa, J. A., & Martínez-Trinidad, J. F. (2013). An Empirical Study of Oversampling and Undersampling for Instance Selection Methods on Imbalance Datasets. In J. Ruiz-Shulcloper & G. Sanniti di Baja (Eds.), Progress in Pattern Recognition, Image Analysis, Computer Vision, and Applications (Vol. 8258, pp. 262–269). Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-41822-8_33 Hooda, N., Bawa, S., & Rana, P. S. (2018). Fraudulent Firm Classification: A Case Study of an External Audit. Applied Artificial Intelligence , 32 (1), 48–64. https://doi.org/10.1080/08839514.2018.1451032 Huan, W., Lin, H., Li, H., Zhou, Y., & Wang, Y. (2020). Anomaly Detection Method Based on Clustering Undersampling and Ensemble Learning. 2020 IEEE 5th Information Technology and Mechatronics Engineering Conference (ITOEC) , 980–984. https://doi.org/10.1109/ITOEC49072.2020.9141897 Huang, G., Wu, L., Ma, X., Zhang, W., Fan, J., Yu, X., Zeng, W., & Zhou, H. (2019). Evaluation of CatBoost method for prediction of reference evapotranspiration in humid regions. Journal of Hydrology , 574 , 1029–1041. https://doi.org/10.1016/j.jhydrol.2019.04.085 Ijaz, M. F., Attique, M., & Son, Y. (2020). Data-Driven Cervical Cancer Prediction Model with Outlier Detection and Over-Sampling Methods. Sensors , 20 (10), 2809. https://doi.org/10.3390/s20102809 Jeni, L. A., Cohn, J. F., & De La Torre, F. (2013). Facing Imbalanced Data—Recommendations for the Use of Performance Metrics. 2013 Humaine Association Conference on Affective Computing and Intelligent Interaction , 245–251. https://doi.org/10.1109/ACII.2013.47 Jo, T., & Japkowicz, N. (2004). Class imbalances versus small disjuncts. ACM SIGKDD Explorations Newsletter , 6 (1), 40–49. https://doi.org/10.1145/1007730.1007737 Kamei, Y., Monden, A., Matsumoto, S., Kakimoto, T., & Matsumoto, K. (2007). The Effects of Over and Under Sampling on Fault-prone Module Detection. First International Symposium on Empirical Software Engineering and Measurement (ESEM 2007) , 196–204. https://doi.org/10.1109/ESEM.2007.28 Keller, A., & Pandey, A. (2021). SMOTE and ENN based XGBoost prediction model for Parkinson’s disease detection. 2021 2nd International Conference on Smart Electronics and Communication (ICOSEC) , 839–846. https://doi.org/10.1109/ICOSEC51865.2021.9591716 Khan, T. M., Xu, S., Khan, Z. G., & Uzair chishti, M. (2021). Implementing Multilabeling, ADASYN, and ReliefF Techniques for Classification of Breast Cancer Diagnostic through Machine Learning: Efficient Computer-Aided Diagnostic System. Journal of Healthcare Engineering , 2021 , 1–15. https://doi.org/10.1155/2021/5577636 Khushi, M., Shaukat, K., Alam, T. M., Hameed, I. A., Uddin, S., Luo, S., Yang, X., & Reyes, M. C. (2021). A Comparative Performance Analysis of Data Resampling Methods on Imbalance Medical Data. IEEE Access , 9 , 109960–109975. https://doi.org/10.1109/ACCESS.2021.3102399 Krayem, A., Yeretzian, A., Faour, G., & Najem, S. (2021). Machine learning for buildings’ characterization and power-law recovery of urban metrics. PLOS ONE , 16 (1), e0246096. https://doi.org/10.1371/journal.pone.0246096 Kurani, A., Doshi, P., Vakharia, A., & Shah, M. (2023). A Comprehensive Comparative Study of Artificial Neural Network (ANN) and Support Vector Machines (SVM) on Stock Forecasting. Annals of Data Science , 10 (1), 183–208. https://doi.org/10.1007/s40745-021-00344-x Kurniawati, Y. E., Permanasari, A. E., & Fauziati, S. (2018). Adaptive Synthetic-Nominal (ADASYN-N) and Adaptive Synthetic-KNN (ADASYN-KNN) for Multiclass Imbalance Learning on Laboratory Test Data. 2018 4th International Conference on Science and Technology (ICST) , 1–6. https://doi.org/10.1109/ICSTC.2018.8528679 Lamari, M., Azizi, N., Hammami, N. E., Boukhamla, A., Cheriguene, S., Dendani, N., & Benzebouchi, N. E. (2021). SMOTE–ENN-Based Data Sampling and Improved Dynamic Ensemble Selection for Imbalanced Medical Data Classification. In F. Saeed, T. Al-Hadhrami, F. Mohammed, & E. Mohammed (Eds.), Advances on Smart and Soft Computing (Vol. 1188, pp. 37–49). Springer Singapore. https://doi.org/10.1007/978-981-15-6048-4_4 Lin, C., Tsai, C.-F., & Lin, W.-C. (2023). Towards hybrid over- and under-sampling combination methods for class imbalanced datasets: An experimental study. Artificial Intelligence Review , 56 (2), 845–863. https://doi.org/10.1007/s10462-022-10186-5 Liu, R., Huang, J., & Zhang, Z. (2023). Tracking disclosure change trajectories for financial fraud detection. Production and Operations Management , 32 (2), 584–602. https://doi.org/10.1111/poms.13888 Lokanan, M. E. (2023). Predicting mobile money transaction fraud using machine learning algorithms. Applied AI Letters , 4 (2), e85. https://doi.org/10.1002/ail2.85 Lokanan, M. E., & Sharma, K. (2022). Fraud prediction using machine learning: The case of investment advisors in Canada. Machine Learning with Applications , 8 , 100269. https://doi.org/10.1016/j.mlwa.2022.100269 Lokanan, M., & Sharma, S. (2024). The use of machine learning algorithms to predict financial statement fraud. The British Accounting Review , 101441. https://doi.org/10.1016/j.bar.2024.101441 Lu, C., Lin, S., Liu, X., & Shi, H. (2020). Telecom Fraud Identification Based on ADASYN and Random Forest. 2020 5th International Conference on Computer and Communication Systems (ICCCS) , 447–452. https://doi.org/10.1109/ICCCS49078.2020.9118521 Maldonado, S., López, J., & Vairetti, C. (2019). An alternative SMOTE oversampling strategy for high-dimensional datasets. Applied Soft Computing , 76 , 380–389. https://doi.org/10.1016/j.asoc.2018.12.024 Mienye, I. D., & Sun, Y. (2023). A Deep Learning Ensemble With Data Resampling for Credit Card Fraud Detection. IEEE Access , 11 , 30628–30638. https://doi.org/10.1109/ACCESS.2023.3262020 More, A. (2016). Survey of resampling techniques for improving classification performance in unbalanced datasets . https://doi.org/10.48550/ARXIV.1608.06048 Moreo, A., Esuli, A., & Sebastiani, F. (2016). Distributional Random Oversampling for Imbalanced Text Classification. Proceedings of the 39th International ACM SIGIR Conference on Research and Development in Information Retrieval , 805–808. https://doi.org/10.1145/2911451.2914722 Mqadi, N. M., Naicker, N., & Adeliyi, T. (2021). Solving Misclassification of the Credit Card Imbalance Problem Using Near Miss. Mathematical Problems in Engineering , 2021 , 1–16. https://doi.org/10.1155/2021/7194728 Munshi, R. M. (2024). Correction: Novel ensemble learning approach with SVM-imputed ADASYN features for enhanced cervical cancer prediction. PLOS ONE , 19 (2), e0298980. https://doi.org/10.1371/journal.pone.0298980 Muranda, C., Ali, A., & Shongwe, T. (2020). Detecting Fraudulent Motor Insurance Claims Using Support Vector Machines with Adaptive Synthetic Sampling Method. 2020 61st International Scientific Conference on Information Technology and Management Science of Riga Technical University (ITMS) , 1–5. https://doi.org/10.1109/ITMS51158.2020.9259322 Naseriparsa, M., & Kashani, M. M. R. (2014). Combination of PCA with SMOTE Resampling to Boost the Prediction Rate in Lung Cancer Dataset . https://doi.org/10.48550/ARXIV.1403.1949 Nayan, N. M., Islam, A., Islam, M. U., Ahmed, E., Hossain, M. M., & Alam, M. Z. (2023). SMOTE Oversampling and Near Miss Undersampling Based Diabetes Diagnosis from Imbalanced Dataset with XAI Visualization. 2023 IEEE Symposium on Computers and Communications (ISCC) , 1–6. https://doi.org/10.1109/ISCC58397.2023.10218281 Nguyen, H. M., Cooper, E. W., & Kamei, K. (2011). Borderline over-sampling for imbalanced data classification. International Journal of Knowledge Engineering and Soft Data Paradigms , 3 (1), 4. https://doi.org/10.1504/IJKESDP.2011.039875 Ning, Q., Zhao, X., & Ma, Z. (2022). A Novel Method for Identification of Glutarylation Sites Combining Borderline-SMOTE With Tomek Links Technique in Imbalanced Data. IEEE/ACM Transactions on Computational Biology and Bioinformatics , 19 (5), 2632–2641. https://doi.org/10.1109/TCBB.2021.3095482 Oladunni, T., Tossou, S., Haile, Y., & Kidane, A. (2021). COVID-19 County Level Severity Classification with Imbalanced Dataset: A NearMiss Under-sampling Approach [Preprint]. Epidemiology. https://doi.org/10.1101/2021.05.21.21257603 Pan, T., Zhao, J., Wu, W., & Yang, J. (2020). Learning imbalanced datasets based on SMOTE and Gaussian distribution. Information Sciences , 512 , 1214–1233. https://doi.org/10.1016/j.ins.2019.10.048 Pereira, R. M., Costa, Y. M. G., & Silla Jr., C. N. (2020). MLTL: A multi-label approach for the Tomek Link undersampling algorithm. Neurocomputing , 383 , 95–105. https://doi.org/10.1016/j.neucom.2019.11.076 Pias, T. S., Su, Y., Tang, X., Wang, H., Faghani, S., & Yao, D. (Daphne). (2023). Enhancing Fairness and Accuracy in Diagnosing Type 2 Diabetes in Young Population [Preprint]. Health Informatics. https://doi.org/10.1101/2023.05.02.23289405 Rajer-Kanduč, K., Zupan, J., & Majcen, N. (2003). Separation of data on the training and test set for modelling: A case study for modelling of five colour properties of a white pigment. Chemometrics and Intelligent Laboratory Systems , 65 (2), 221–229. https://doi.org/10.1016/S0169-7439(02)00110-7 Rao, C., Xu, Y., Xiao, X., Hu, F., & Goh, M. (2024). Imbalanced customer churn classification using a new multi-strategy collaborative processing method. Expert Systems with Applications , 247 , 123251. https://doi.org/10.1016/j.eswa.2024.123251 Rubaidi, Z. S., Ammar, B. B., & Aouicha, M. B. (2022). Fraud Detection Using Large-scale Imbalance Dataset. International Journal on Artificial Intelligence Tools , 31 (08), 2250037. https://doi.org/10.1142/S0218213022500373 Santos, M. S., Soares, J. P., Abreu, P. H., Araujo, H., & Santos, J. (2018). Cross-Validation for Imbalanced Datasets: Avoiding Overoptimistic and Overfitting Approaches [Research Frontier]. IEEE Computational Intelligence Magazine , 13 (4), 59–76. https://doi.org/10.1109/MCI.2018.2866730 Saripuddin, M., Suliman, A., Syarmila Sameon, S., & Jorgensen, B. N. (2021). Random Undersampling on Imbalance Time Series Data for Anomaly Detection. 2021 The 4th International Conference on Machine Learning and Machine Intelligence , 151–156. https://doi.org/10.1145/3490725.3490748 Silva, B., Silveira, R., Silva Neto, M., Cortez, P., & Gomes, D. (2021). A comparative analysis of undersampling techniques for network intrusion detection systems design. Journal of Communication and Information Systems , 36 (1), 31–43. https://doi.org/10.14209/jcis.2021.3 Singh, A., Ranjan, R. K., & Tiwari, A. (2022). Credit Card Fraud Detection under Extreme Imbalanced Data: A Comparative Study of Data-level Algorithms. Journal of Experimental & Theoretical Artificial Intelligence , 34 (4), 571–598. https://doi.org/10.1080/0952813X.2021.1907795 Sisodia, D. S., Reddy, N. K., & Bhandari, S. (2017). Performance evaluation of class balancing techniques for credit card fraud detection. 2017 IEEE International Conference on Power, Control, Signals and Instrumentation Engineering (ICPCSI) , 2747–2752. https://doi.org/10.1109/ICPCSI.2017.8392219 Srinilta, C., & Kanharattanachai, S. (2021). Application of Natural Neighbor-based Algorithm on Oversampling SMOTE Algorithms. 2021 7th International Conference on Engineering, Applied Sciences and Technology (ICEAST) , 217–220. https://doi.org/10.1109/ICEAST52143.2021.9426310 Srinivasan, S., Vallikannu, A. L., Manoharan, L., Deepthi, K., & Aravind Yadav, B. (2024). Identification of the Best Combination of Oversampling Technique and Machine Learning Algorithm for Credit Card Fraud Detection. In I. J. Jacob, S. Piramuthu, & P. Falkowski-Gilski (Eds.), Data Intelligence and Cognitive Informatics (pp. 557–571). Springer Nature Singapore. https://doi.org/10.1007/978-981-99-7962-2_41 Subudhi, S., & Panigrahi, S. (2018). Effect of Class Imbalanceness in Detecting Automobile Insurance Fraud. 2018 2nd International Conference on Data Science and Business Analytics (ICDSBA) , 528–531. https://doi.org/10.1109/ICDSBA.2018.00104 Sujitha, R., & Paramasivan, B. (2023). Optimal progressive classification study using SMOTE-SVM for stages of lung disease. Automatika , 64 (4), 807–814. https://doi.org/10.1080/00051144.2023.2218167 Sun, B., & Chen, H. (2021). A Survey of k Nearest Neighbor Algorithms for Solving the Class Imbalanced Problem. Wireless Communications and Mobile Computing , 2021 , 1–12. https://doi.org/10.1155/2021/5520990 Sun, T., & Vasarhelyi, M. A. (2018). Predicting credit card delinquencies: An application of deep neural networks. Intelligent Systems in Accounting, Finance and Management , 25 (4), 174–189. https://doi.org/10.1002/isaf.1437 Tanimu, J. J., Hamada, M., Hassan, M., Kakudi, H., & Abiodun, J. O. (2022). A Machine Learning Method for Classification of Cervical Cancer. Electronics , 11 (3), 463. https://doi.org/10.3390/electronics11030463 Tariq, M. A., Sargano, A. B., Iftikhar, M. A., & Habib, Z. (2023). Comparing Different Oversampling Methods in Predicting Multi-Class Educational Datasets Using Machine Learning Techniques. Cybernetics and Information Technologies , 23 (4), 199–212. https://doi.org/10.2478/cait-2023-0044 Tiwari, A. K., Nath, A., Subbiah, K., & Shukla, K. K. (2015). Effect of varying degree of resampling on prediction accuracy for observed peptide count in protein mass spectrometry data. 2015 11th International Conference on Natural Computation (ICNC) , 691–695. https://doi.org/10.1109/ICNC.2015.7378074 Tiwari, S., Wee, H. M., & Daryanto, Y. (2018). Big data analytics in supply chain management between 2010 and 2016: Insights to industries. Computers & Industrial Engineering , 115 , 319–330. https://doi.org/10.1016/j.cie.2017.11.017 Vuttipittayamongkol, P., Elyan, E., & Petrovski, A. (2021). On the class overlap problem in imbalanced data classification. Knowledge-Based Systems , 212 , 106631. https://doi.org/10.1016/j.knosys.2020.106631 Wang, K., Tian, J., Zheng, C., Yang, H., Ren, J., Li, C., Han, Q., & Zhang, Y. (2021). Improving Risk Identification of Adverse Outcomes in Chronic Heart Failure Using SMOTE+ENN and Machine Learning. Risk Management and Healthcare Policy , Volume 14 , 2453–2463. https://doi.org/10.2147/RMHP.S310295 Wang, Z., Wu, C., Zheng, K., Niu, X., & Wang, X. (2019). SMOTETomek-Based Resampling for Personality Recognition. IEEE Access , 7 , 129678–129689. https://doi.org/10.1109/ACCESS.2019.2940061 Xie, X., Liu, H., Zeng, S., Lin, L., & Li, W. (2021). A novel progressively undersampling method based on the density peaks sequence for imbalanced data. Knowledge-Based Systems , 213 , 106689. https://doi.org/10.1016/j.knosys.2020.106689 Xu, Z., Shen, D., Nie, T., & Kou, Y. (2020). A hybrid sampling algorithm combining M-SMOTE and ENN based on Random forest for medical imbalanced data. Journal of Biomedical Informatics , 107 , 103465. https://doi.org/10.1016/j.jbi.2020.103465 Xu, Z., Shen, D., Nie, T., Kou, Y., Yin, N., & Han, X. (2021). A cluster-based oversampling algorithm combining SMOTE and k-means for imbalanced medical data. Information Sciences , 572 , 574–589. https://doi.org/10.1016/j.ins.2021.02.056 Xuan, S., Liu, G., Li, Z., Zheng, L., Wang, S., & Jiang, C. (2018). Random forest for credit card fraud detection. 2018 IEEE 15th International Conference on Networking, Sensing and Control (ICNSC) , 1–6. https://doi.org/10.1109/ICNSC.2018.8361343 Yang, F., Wang, K., Sun, L., Zhai, M., Song, J., & Wang, H. (2022). A hybrid sampling algorithm combining synthetic minority over-sampling technique and edited nearest neighbor for missed abortion diagnosis. BMC Medical Informatics and Decision Making , 22 (1), 344. https://doi.org/10.1186/s12911-022-02075-2 Yang, Y., Yang, X., Tang, W., & Li, L. (2023). A Undersampling-DoppelGANger based Data Generation Method for Unbalanced BGP Data. 2023 IEEE 9th International Conference on Cloud Computing and Intelligent Systems (CCIS) , 100–105. https://doi.org/10.1109/CCIS59572.2023.10263221 Yap, B. W., Rani, K. A., Rahman, H. A. A., Fong, S., Khairudin, Z., & Abdullah, N. N. (2014). An Application of Oversampling, Undersampling, Bagging and Boosting in Handling Imbalanced Datasets. In T. Herawan, M. M. Deris, & J. Abawajy (Eds.), Proceedings of the First International Conference on Advanced Data and Information Engineering (DaEng-2013) (Vol. 285, pp. 13–22). Springer Singapore. https://doi.org/10.1007/978-981-4585-18-7_2 Zeng, M., Zou, B., Wei, F., Liu, X., & Wang, L. (2016). Effective prediction of three common diseases by combining SMOTE with Tomek links technique for imbalanced medical data. 2016 IEEE International Conference of Online Analysis and Computing Science (ICOACS) , 225–228. https://doi.org/10.1109/ICOACS.2016.7563084 Zhang, H., & Li, M. (2014). RWO-Sampling: A random walk over-sampling approach to imbalanced data classification. Information Fusion , 20 , 99–116. https://doi.org/10.1016/j.inffus.2013.12.003 Zhao, H., Chen, X., Nguyen, T., Huang, J. Z., Williams, G., & Chen, H. (2016). Stratified Over-Sampling Bagging Method for Random Forests on Imbalanced Data. In M. Chau, G. A. Wang, & H. Chen (Eds.), Intelligence and Security Informatics (Vol. 9650, pp. 63–72). Springer International Publishing. https://doi.org/10.1007/978-3-319-31863-9_5 Zhou, L. (2013). Performance of corporate bankruptcy prediction models on imbalanced dataset: The effect of sampling methods. Knowledge-Based Systems , 41 , 16–25. https://doi.org/10.1016/j.knosys.2012.12.007 Zhu, T., Lin, Y., & Liu, Y. (2017). Synthetic minority oversampling technique for multiclass imbalance problems. Pattern Recognition , 72 , 327–340. https://doi.org/10.1016/j.patcog.2017.07.024 Zuech, R., Hancock, J., & Khoshgoftaar, T. M. (2021). Detecting web attacks using random undersampling and ensemble learners. Journal of Big Data , 8 (1), 75. https://doi.org/10.1186/s40537-021-00460-8 Footnotes Yeh,I-Cheng. (2016). Default of credit card clients. UCI Machine Learning Repository. https://doi.org/10.24432/C55S3H . Additional Declarations The authors declare no competing interests. Cite Share Download PDF Status: Posted Version 1 posted You are reading this latest preprint version Research Square lets you share your work early, gain feedback from the community, and start making changes to your manuscript prior to peer review in a journal. As a division of Research Square Company, we’re committed to making research communication faster, fairer, and more useful. We do this by developing innovative software and high quality services for the global research community. Our growing team is made up of researchers and industry professionals working together to solve the most critical problems facing scientific publishing. Also discoverable on Platform About Our Team In Review Editorial Policies Advisory Board Help Center Resources Author Services Accessibility API Access RSS feed Manage Cookie Preferences © Research Square 2026 | ISSN 2693-5015 (online) Privacy Policy Terms of Service Do Not Sell My Personal Information {"props":{"pageProps":{"initialData":{"identity":"rs-6895500","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":471346761,"identity":"928c9d24-9139-4f53-b2ab-ca554849aff0","order_by":0,"name":"Mark Lokanan","email":"data:image/png;base64,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","orcid":"","institution":"Royal Roads University","correspondingAuthor":true,"prefix":"","firstName":"Mark","middleName":"","lastName":"Lokanan","suffix":""}],"badges":[],"createdAt":"2025-06-14 19:32:11","currentVersionCode":1,"declarations":{"humanSubjects":false,"vertebrateSubjects":false,"conflictsOfInterestStatement":false,"humanSubjectEthicalGuidelines":false,"humanSubjectConsent":false,"humanSubjectClinicalTrial":false,"humanSubjectCaseReport":false,"vertebrateSubjectEthicalGuidelines":false},"doi":"10.21203/rs.3.rs-6895500/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-6895500/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":84868939,"identity":"6b5cf27a-38aa-4367-bbd6-f77a5f7eec21","added_by":"auto","created_at":"2025-06-18 08:45:54","extension":"jpeg","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":78608,"visible":true,"origin":"","legend":"\u003cp\u003eUndersampling Minority Class\u003c/p\u003e","description":"","filename":"floatimage1.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6895500/v1/cb3baff19bbd8f34b9d8ee05.jpeg"},{"id":84868940,"identity":"c9dbda03-9605-4ed8-85e1-b0fff5a6953b","added_by":"auto","created_at":"2025-06-18 08:45:54","extension":"jpeg","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":79224,"visible":true,"origin":"","legend":"\u003cp\u003eOversampling Majority Class\u003c/p\u003e","description":"","filename":"floatimage2.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6895500/v1/631997f1ecef3cff1f7018c6.jpeg"},{"id":84870313,"identity":"824597af-82c3-4bf3-8996-695dab806479","added_by":"auto","created_at":"2025-06-18 08:53:54","extension":"jpeg","order_by":3,"title":"Figure 3","display":"","copyAsset":false,"role":"figure","size":157349,"visible":true,"origin":"","legend":"\u003cp\u003eSMOTE Oversampling\u003c/p\u003e","description":"","filename":"floatimage3.jpeg","url":"https://assets-eu.researchsquare.com/files/rs-6895500/v1/08d869b6488ec51ed66b6ef3.jpeg"},{"id":84868945,"identity":"b063482c-d2a0-4022-8a80-7ea55ef07d26","added_by":"auto","created_at":"2025-06-18 08:45:54","extension":"png","order_by":4,"title":"Figure 4","display":"","copyAsset":false,"role":"figure","size":153009,"visible":true,"origin":"","legend":"\u003cp\u003eDeep Adaptive Resampling Technique\u003c/p\u003e","description":"","filename":"floatimage4.png","url":"https://assets-eu.researchsquare.com/files/rs-6895500/v1/54f3a17eaa170cd3b20e3b03.png"},{"id":84870988,"identity":"077f4a2d-9c96-45d2-a05f-8f7d5a48d186","added_by":"auto","created_at":"2025-06-18 09:01:57","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":2939822,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-6895500/v1/b32f4356-31be-43f0-abf9-ed2d423aa1f5.pdf"}],"financialInterests":"The authors declare no competing interests.","formattedTitle":"\u003cp\u003eIntroducing DART: A Novel Deep Adaptive Upsampling Technique for Handling Class Imbalance\u003c/p\u003e","fulltext":[{"header":"1 Introduction","content":"\u003cp\u003eThe success and accuracy of machine learning models depend heavily on the quality and distribution of the data. One of the most persistent challenges in this regard is class imbalance, which can significantly distort model performance by biasing predictions toward the majority class. Imbalanced data occurs when one class is heavily underrepresented compared to the other, leading to biased predictions and poor generalization, especially for the minority class. Data imbalance is frequently encountered in real-world scenarios such as credit and loan default prediction, fraud detection, and disease diagnosis, where the event of interest (e.g., default, fraud, or illness) occurs far less frequently than the non-event.\u003c/p\u003e \u003cp\u003eImbalanced datasets present numerous challenges for machine learning algorithms (Araf et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Cinaroglu, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). The skewed distribution of classes can introduce biases during model training, favouring the majority class and resulting in diminished performance for the minority class. The dominance of the majority class within the dataset often leads algorithms to prioritize predicting this class, potentially overlooking the minority class, resulting in biased classifications and poor generalization. Imbalanced classification scenarios, also known as rare event modelling, arise when the target variable exhibits a significant imbalance, with the minority class representing the rare events. In such cases, the model's tendency to learn predominantly from the majority class can make predicting the minority class particularly difficult (Araf et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Khushi et al., \u003cspan citationid=\"CR54\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Naseriparsa \u0026amp; Kashani, \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Consequently, machine learning algorithms may struggle to construct accurate models (Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e), leading to challenges in evaluation metrics that could fall into the \"metric trap\" and result in inaccurate results (Jeni et al., \u003cspan citationid=\"CR49\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Yap et al., \u003cspan citationid=\"CR107\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWhen handling imbalanced data, many machine learning algorithms tend to excel at accurately predicting the majority class while struggling with the minority class (Zhou, \u003cspan citationid=\"CR111\" class=\"CitationRef\"\u003e2013\u003c/span\u003e). In datasets with class imbalances, classification algorithms often prioritize predicting the majority class without adequately analyzing the minority class (T. Sun \u0026amp; Vasarhelyi, \u003cspan citationid=\"CR93\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Consequently, this can lead to consistently high-performance accuracy scores that are not necessarily reliable (An \u0026amp; Suh, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Hooda et al., \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Lin et al., \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Relying solely on raw performance accuracy may not be ideal for evaluating classification models, especially in the context of imbalanced datasets (Araf et al., \u003cspan citationid=\"CR10\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Hooda et al., \u003cspan citationid=\"CR45\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). This outcome is because accuracy only considers true positive/negative observations and ignores false positive and negative observations. While alternative metrics such as precision, recall, F1-score, and Area Under Receiver Operating Curve (ROC) Curve (AUROC) offer more robust evaluation options for imbalanced datasets, they still tend to favour the majority class (Fujiwara et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Hasanin et al., \u003cspan citationid=\"CR43\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eThe purpose of this study is to tackle class imbalance in predictive modeling, a problem that can severely distort machine learning performance by favoring the majority class. Rather than focusing on general algorithmic adjustments or cost-sensitive learning, this paper concentrates on resampling strategies\u0026mdash;both undersampling and oversampling\u0026mdash;as effective methods for correcting skewed class distributions. To achieve this purpose, the study is guided by two main objectives:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eTo systematically evaluate and compare the effectiveness of established resampling techniques across a diverse set of machine learning classifiers using both a credit and a loan default dataset.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eTo introduce and assess the performance of a novel resampling method, DART (Deep Adaptive Resampling Technique), by applying it to both the credit default dataset and a secondary loan default dataset, thereby testing its robustness and generalizability.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eIn addressing these objectives, the paper makes several important contributions to the literature on handling imbalanced datasets. First, while prior studies often focus on a single undersampling or oversampling method applied to one dataset, the present study conducts a comprehensive evaluation of multiple resampling techniques across two distinct datasets\u0026mdash;credit default and loan default. By assessing the performance of each method across a range of machine learning algorithms and datasets, the study provides a broader and more robust understanding of which techniques yield the most effective results in different predictive contexts. Second, the comparative analysis of undersampling and oversampling methods generates practical scientific insights that bridge the gap between theoretical advancements and real-world applications. Highlighting the trade-offs, strengths, and limitations of each technique empowers researchers and practitioners to make more informed methodological choices when working with imbalanced data. Third, a novel resampling technique\u0026mdash;DART\u0026mdash;is introduced and applied to both datasets. DART dynamically adjusts the sampling process based on model feedback, offering a flexible and adaptive solution to the class imbalance problem. Its integration into the modeling pipeline demonstrates its potential to improve classification performance, particularly in the domain of financial risk prediction.\u003c/p\u003e \u003cp\u003eThe structure of this paper is organized as follows. Section \u003cspan refid=\"Sec2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents a review of the literature on undersampling techniques, highlighting their theoretical foundations and applications. Section \u003cspan refid=\"Sec8\" class=\"InternalRef\"\u003e3\u003c/span\u003e examines the body of research on oversampling methods, with particular emphasis on SMOTE-based and hybrid approaches. Section 4 critically evaluates the advantages and drawbacks of both undersampling and oversampling strategies, especially in the context of model performance on imbalanced datasets. Section \u003cspan refid=\"Sec20\" class=\"InternalRef\"\u003e5\u003c/span\u003e outlines the methodology, detailing the datasets, resampling techniques, and classification algorithms employed. Section \u003cspan refid=\"Sec31\" class=\"InternalRef\"\u003e6\u003c/span\u003e provides a comprehensive analysis of the empirical results, comparing the effectiveness of the various resampling approaches across multiple metrics. Finally, Section \u003cspan refid=\"Sec39\" class=\"InternalRef\"\u003e7\u003c/span\u003e offers a discussion and conclusion, synthesizing the key findings and outlining their implications for future research and practical applications.\u003c/p\u003e"},{"header":"2 Literature Review: Resampling Imbalanced Datasets","content":"\u003cp\u003eA well-established method for addressing the class imbalance problem is resampling the data (Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e; Ghorbani \u0026amp; Ghousi, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Lin et al., \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). The objective of resampling is to ensure that the samples used in the model closely resemble the population they originate from, facilitating accurate inferences about the true population from the sample (Rajer-Kanduč et al., \u003cspan citationid=\"CR80\" class=\"CitationRef\"\u003e2003\u003c/span\u003e). Resampling encompasses two main approaches: undersampling, which involves removing observations from the majority class to align with the minority class, and oversampling, which entails adding more observations to the minority class (An \u0026amp; Suh, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Cinaroglu, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Xuan et al., \u003cspan citationid=\"CR104\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Both undersampling and oversampling techniques aim to adjust the class ratio in an imbalanced dataset for modelling purposes.\u003c/p\u003e \u003cp\u003eThe subsequent section reviews the literature on the undersampling and oversampling techniques that have gained prominence over the years. Special attention will be given to their effectiveness in addressing class imbalances, their application in various domains, and their impact on model performance and generalization. The literature to follow is illustrative rather than exhaustive. The intention is to explore the use and applicability of undersampling and overs-sampling techniques in machine learning applications that have been in ascendency in the literature.\u003c/p\u003e \u003cdiv id=\"Sec3\" class=\"Section2\"\u003e \u003ch2\u003e2.1 Undersampling Majority Class\u003c/h2\u003e \u003cp\u003eUndersampling is a technique where the number of observations from the majority class is reduced to match the minority class (An \u0026amp; Suh, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Dal Pozzolo et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Zuech et al., \u003cspan citationid=\"CR113\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e1\u003c/span\u003e presents a diagrammatic illustration of the undersampling technique, where there are 12,000 observations of the majority class and 2,000 observations of the minority class. A model built with this sample distribution will be biased towards the majority class because the sample is more likely to be fed into the algorithm multiple times compared to the minority class (An \u0026amp; Suh, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Hernandez et al., \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Santos et al., \u003cspan citationid=\"CR83\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e1\u003c/span\u003e, to create a balanced dataset, the data is resampled with random samples of 2,000 observations from the majority class to match the 2,000 observations of the minority class (An \u0026amp; Suh, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eUndersampling is often favoured due to the presence of a sufficient minority class, allowing for a balanced 50\u0026ndash;50 split while maintaining a representative sample. This approach aids data mining algorithms in effectively analyzing the data within manageable limits (Dal Pozzolo et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Fujiwara et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). With undersampling, the majority of class observations are discarded at random until there is a more balanced distribution of the data. Balancing out the minority and majority classes allows the classifier to weigh both classes equally and produce more representative results (Dal Pozzolo et al., \u003cspan citationid=\"CR29\" class=\"CitationRef\"\u003e2014\u003c/span\u003e; Hernandez et al., \u003cspan citationid=\"CR44\" class=\"CitationRef\"\u003e2013\u003c/span\u003e; Yap et al., \u003cspan citationid=\"CR107\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Several undersampling techniques commonly utilized in the literature include random undersampling, Tomek links, NearMiss, and Condensed Nearest Neighbour (CNN). Below, we examine the existing literature assembled to date on these undersampling techniques.\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec4\" class=\"Section2\"\u003e \u003ch2\u003e2.2 Random undersampling\u003c/h2\u003e \u003cp\u003eRandom undersampling (RUS) is a technique used in machine learning to address the issue of imbalanced datasets, where the majority class significantly outnumbers the minority class (Kamei et al., \u003cspan citationid=\"CR51\" class=\"CitationRef\"\u003e2007\u003c/span\u003e; Saripuddin et al., \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Zuech et al., \u003cspan citationid=\"CR113\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). This approach involves randomly selecting a subset of samples from the majority class and removing them from the dataset to achieve a more balanced class distribution. Practitioners can implement RUS by specifying either a desired ratio of majority to minority class samples or a fixed number of samples to be removed from the majority class. RUS is simple to implement and computationally time- and resource-efficient (Zuech et al., \u003cspan citationid=\"CR113\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eRecent studies on RUS in machine learning span across various domains, each exploring the technique's effectiveness in addressing class imbalance problems in specific domains. Some researchers have employed RUS to improve classification performance in anomaly detection (Huan et al., \u003cspan citationid=\"CR46\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Saripuddin et al., \u003cspan citationid=\"CR84\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Y. Yang et al., \u003cspan citationid=\"CR106\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). These studies underscore the versatility of RUS in different domains to detect anomalies and improve the performance of machine learning models. Others have utilized RUS to address the challenges of class imbalance in cybersecurity datasets (Bagui \u0026amp; Li, \u003cspan citationid=\"CR12\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Silva et al., \u003cspan citationid=\"CR85\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Zuech et al., \u003cspan citationid=\"CR113\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). By employing RUS, these researchers were able to enhance detection accuracy for cybersecurity threats by making the models more reliable and efficient in identifying potential threats in vast amounts of normal traffic data. Another area where RUS has been utilized with very good performance is in predictive modeling and disease diagnosis in the healthcare industry (Bauder \u0026amp; Khoshgoftaar, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Researchers have applied RUS to Medicare fraud detection, showing that it significantly boosts classification accuracy (J. Hancock et al., \u003cspan citationid=\"CR41\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Pias et al. (\u003cspan citationid=\"CR79\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) explore the use of RUS to balance datasets in predicting diabetes and prediabetes in patients. The authors found that RUS achieved robust results and enhanced the fairness and performance of machine learning models (Pias et al., \u003cspan citationid=\"CR79\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Others have noted that a 50:50 RUS does not produce the best Medicare fraud detection results; rather, the authors found that a 90:10 class distribution offers the best detection (Bauder \u0026amp; Khoshgoftaar, \u003cspan citationid=\"CR19\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). These findings suggest that RUS may not be the best technique for fraud prediction in the healthcare industry. That aside, these studies collectively underscore the significance of RUS in enhancing the performance of machine learning models in a variety of domains.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec5\" class=\"Section2\"\u003e \u003ch2\u003e2.3 Tomek links\u003c/h2\u003e \u003cp\u003eThe Tomek links undersampling method is used to balance imbalanced datasets by removing observations located between two different classes. The basic idea behind Tomek links is to identify pairs of observations, one from the majority class and one from the minority class, that are closest to each other but belong to different classes (Alamri \u0026amp; Ykhlef, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Devi et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). These pairs are called Tomek links (Devi et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e, p. 3). Tomek links works by eliminating instances belonging to the majority class from Tomek linkages. Removing the majority class can effectively enhance the separation between classes and boost the efficiency of classification systems. Tomek links focuses on the ambiguous cases that are prone to causing misclassification and eliminates them to provide a more distinct decision boundary between the classes. Tomek links aims to improve the overall balance of the dataset while preserving the minority class instances that are farthest from the majority class.\u003c/p\u003e \u003cp\u003eResearch in Tomek links to address class imbalance spans a variety of domains and demonstrates the versatility and effectiveness of the technique to enhance the performance of machine learning models in different applications. Some studies propose an approach combining one-class SVM for anomaly detection with adapted Tomek links pairs to eliminate redundant and overlapping cases and address imbalance data issues (Basit et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Devi et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Vuttipittayamongkol et al., \u003cspan citationid=\"CR98\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eInterestingly, the findings suggest that class overlap has a more detrimental effect on performance than class imbalance. That said, there is evidence to show that Tomek links has been very effective in addressing imbalances and overlapping issues in medical datasets (Basit et al., \u003cspan citationid=\"CR17\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Others have used Tomek links in the area of multi-label classification to remove boundary and noise samples from datasets (Ai-jun \u0026amp; Peng, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Pereira et al., \u003cspan citationid=\"CR78\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Like Devi et al. (2019), these studies address imbalance by selectively removing overlapping samples between classes to improve the performance of multi-label classification (Ai-jun \u0026amp; Peng, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Pereira et al., \u003cspan citationid=\"CR78\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Others have used Tomek links in bioinformatics and medical diagnostics to improve machine learning model performance on imbalanced datasets (Ning et al., \u003cspan citationid=\"CR75\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Zeng et al., \u003cspan citationid=\"CR108\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). The authors note that Tomek links, in combination with the synthetic minority oversampling technique (SMOTE), enhances classification accuracy across different metrics and demonstrates the benefits of combining these resampling techniques for medical data classification (Zeng et al., \u003cspan citationid=\"CR108\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). These studies illustrate a growing interest in using Tomek links to address the intertwined challenges of class imbalance and overlapping issues across different applications.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec6\" class=\"Section2\"\u003e \u003ch2\u003e2.4 NearMiss\u003c/h2\u003e \u003cp\u003eNearMiss is an undersampling method that balances imbalanced datasets by removing observations from the majority class closest to the minority class. NearMiss is a k-nearest neighbor approach that balances the class distribution by choosing instances based on the distance between the majority class and the minority class (Mqadi et al., \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, pp. 3\u0026ndash;4; Oladunni et al., \u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, p. 3). When the distance between the majority and minority classes is too close, NearMiss removes instances of the majority class in order to increase the distance between them (Ha \u0026amp; Lee, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2016\u003c/span\u003e, p. 2). To find the closest instances of the majority class, there are three types of NearMiss techniques: NearMiss-1, NearMiss-2, and NearMiss-3. NearMiss-1 selects instances from the majority class whose individual distance to the three closest instances to the minority class is the smallest; NearMiss-2 selects instances from the majority class whose individual distance to the three farthest instances to the minority class is the smallest; and NearMiss-3 selects instances from the majority class for which each instance in the minority class has the closest distance (Ha \u0026amp; Lee, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2016\u003c/span\u003e, p. 2). The commonality of the NearMiss family is that they select the instances in the majority class that are closest to the minority class to better learn the decision boundary in the data (p. 2).\u003c/p\u003e \u003cp\u003eMost of the literature on NearMiss undersampling is sparse, with few studies exploring its effectiveness across various domains. Bao et al. (\u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) proposed boosted near-miss undersampling on SVM (BNU-SVMs) ensembles for concept detection in large-scale imbalanced datasets. The authors discovered that BNU-SVMs may effectively manage large-scale imbalanced datasets by balancing and reducing the training dataset through undersampling (L. Bao et al., \u003cspan citationid=\"CR14\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). The classifier's performance is enhanced by integrating multiple classifiers. Other studies have employed NearMiss undersampling to address imbalances in financial crime datasets (Mqadi et al., \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Rubaidi et al., \u003cspan citationid=\"CR82\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The studies found that machine learning algorithms performed very well using the NearMiss undersampling technique. NearMiss undersampling has also been used in detecting insider threats, specifically data leakage by malicious insiders prior to leaving an organization (Alsowail, \u003cspan citationid=\"CR8\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). The author found that NearMiss undersampling achieved enhanced performance in improving the detection of insider data leakage. Further research in healthcare has found NearMiss to be a promising undersampling technique in bioinformatics and predicting disease with high performance accuracy (Alamsyah et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Nayan et al., \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec7\" class=\"Section2\"\u003e \u003ch2\u003e2.5 Condensed Nearest Neighbor (CNN)\u003c/h2\u003e \u003cp\u003eThe Condensed Nearest Neighbor (CNN) undersampling algorithm operates by systematically removing redundant instances from the majority class that are correctly classified by their nearest neighbor in the minority class while preserving essential information in the dataset. The algorithm follows two main steps:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eInitially, a random subset of the majority class samples is selected to form the \"condensed set.\"\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eThe algorithm then iteratively eliminates samples from the majority class that the condensed set's nearest neighbor correctly classified. This procedure keeps iterating until every sample in the condensed set receives the wrong classification from their nearest neighbor in the majority class.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eThe CNN algorithm employs the 1-NN rule, which dictates that all minority class instances are assigned to set S, one majority class instance is placed in set S, and the remaining majority class instances are allocated to the set (Chaplot et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019\u003c/span\u003e, p. 95). Each sample from set \u003cem\u003eC\u003c/em\u003e is individually assessed using the 1-NN rule (p. 95). If correctly classified, the sample is discarded; otherwise, it is moved to set \u003cem\u003eS\u003c/em\u003e (p. 95). This process repeats for all instances until all the instances have been evaluated using the 1-NN rule (p. 95). The underlying concept behind CNN is that instances correctly classified by their nearest neighbor are considered uninformative for establishing the decision boundary between classes and can be safely removed without compromising classification performance (Batista et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2004\u003c/span\u003e; Chaplot et al., \u003cspan citationid=\"CR24\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Devi et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWhile CNN is limited in application, there are a few studies that provide evidence of its effectiveness in addressing class imbalance. In their study, Bansal and Jain employed CNN as a method to balance the number of instances between two classes (Bansal \u0026amp; Jain, \u003cspan citationid=\"CR13\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). They achieve this by undersampling the majority class according to specific criteria and finding that CNN was the best performer among others. Batista et al. tested CNN along with other techniques across thirteen different datasets and found that class imbalance does not hinder the performance of learning systems (Batista et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). The problems appear to be rooted in learning with too few minority classes in the presence of other heterogeneous factors such as class overlapping (p. 20). In another study, Xie et al. \u003cspan citationid=\"CR101\" class=\"CitationRef\"\u003e2021\u003c/span\u003e employed CNN and other undersampling techniques on 40 public benchmark datasets. The study discovered that CNNs eliminate noisy or boundary occurrences from the majority class, which can be advantageous for learning models dealing with imbalanced data (Xie et al., \u003cspan citationid=\"CR101\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Nevertheless, even when using undersampling techniques, a significant disparity between the majority class and the minority class might still negatively impact the performance of the learning process (Xie et al., \u003cspan citationid=\"CR101\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, p. 8).\u003c/p\u003e \u003cp\u003eEach undersampling strategy has its own distinct advantages in mitigating class imbalances in machine learning applications. RUS has exhibited adaptability and enhanced effectiveness in multiple areas, such as anomaly detection, cybersecurity, and healthcare. Its implementation has been recognized for its ability to improve machine learning models. However, Tomek links specifically aims to address ambiguous observations between different classes. Tomek links has shown potential in the fields of medical diagnostics and multi-label classification, where it has been successfully utilized to enhance model performance by getting rid of irrelevant information and resolving overlaps in the data. NearMiss effectively enhances the distinction between classes by deliberately eliminating instances of the majority class that are closest to the minority class. NearMiss has been employed to enhance machine learning models in the domains of financial crime detection and healthcare. Although CNN and other undersampling approaches provide potential for addressing class imbalance problems in machine learning, the continued existence of substantial differences between the majority and minority classes highlights the persistent difficulty in reaching optimal performance across varied datasets. While these undersampling strategies offer unique advantages in mitigating class imbalance across diverse domains, challenges persist in achieving optimal performance due to significant disparities between the majority and minority classes. Oversampling techniques provide an opportunity to obtain insights on how these methods have been used to address class imbalances in datasets.\u003c/p\u003e \u003c/div\u003e"},{"header":"3 Oversampling Minority Class","content":"\u003cp\u003eRandom oversampling of the minority class is done by duplicating the instances and their representation in the dataset (More, \u003cspan citationid=\"CR67\" class=\"CitationRef\"\u003e2016\u003c/span\u003e, pp. 5\u0026ndash;8). Referring back to the fraud (2%) and non-fraud (98%) examples above, the majority of the data will be skewed towards the no-fraud class. In this case, because there is only 2% of the fraud class, the model will mostly train on the 98% of the no-fraud class. One way of oversampling is to generate new instances for the minority class by sampling with replacement (An \u0026amp; Suh, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e). Figure\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e presents a diagrammatic illustration of random oversampling. Note from Fig.\u0026nbsp;\u003cspan refid=\"Fig2\" class=\"InternalRef\"\u003e2\u003c/span\u003e that 2000 transactions were duplicated six times to balance the data. Oversampling is preferred when there is an abundance of data for the majority class and rare events for the minority class (Elreedy \u0026amp; Atiya, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Yap et al., \u003cspan citationid=\"CR107\" class=\"CitationRef\"\u003e2014\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cdiv id=\"Sec9\" class=\"Section2\"\u003e \u003ch2\u003e3.1 Random Oversampling\u003c/h2\u003e \u003cp\u003eRandom oversampling (ROS) is a method used in handling imbalanced datasets, where instances from the minority class are duplicated randomly to augment the dataset until it reaches the desired ratio or balance with the majority class (Elreedy \u0026amp; Atiya, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; Nayan et al., \u003cspan citationid=\"CR73\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Unlike other techniques, ROS does not consider the similarity and characteristics of the data points and simply duplicates the instances without considering the relevant features in the dataset. ROS typically involves the following steps:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eDetermining the number of samples in the minority class.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eRandomly selecting instances from the minority class and duplicating them to increase the overall count of minority class samples.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003eContinuing the duplication process until the number of minority class samples matches that of the majority class.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eWhile ROS can help address class imbalance, it can also lead to overfitting, especially when the same observations are replicated through multiple iterations in the dataset.\u003c/p\u003e \u003cp\u003eStudies that employ ROS have added some other unique techniques to enhance model performance. In one study, the issue of imbalanced data in binary text classification is tackled through the introduction of distributional random oversampling (Moreo et al., \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). This method utilizes the distributional hypothesis, which posits that the meaning of a feature is shaped by its distribution across extensive data corpora, in order to create synthetic minority-class documents. The results suggest that the distributed random oversampling methods enhance the accuracy of classification algorithms by creating balanced datasets. Others have used ROS to address imbalances in multilabel datasets with very good results across different classification measures (Charte et al., \u003cspan citationid=\"CR25\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). In another variation of ROS, Zhao et al. (\u003cspan citationid=\"CR110\" class=\"CitationRef\"\u003e2016\u003c/span\u003e) address the class imbalance through stratified random oversampling. The study found that the proposed stratified oversampling method effectively addresses the challenge of imbalanced data by generating balanced and diverse training datasets (Zhao et al., \u003cspan citationid=\"CR110\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Others have introduced the Random OverSampling approach to balancing the minority and majority classes by creating synthetic samples by randomly walking from the real data (Zhang \u0026amp; Li, \u003cspan citationid=\"CR109\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). The study found that random walk oversampling statistically performs much better than alternative methods on imbalanced datasets when implementing common baseline algorithms (p. 99).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec10\" class=\"Section2\"\u003e \u003ch2\u003e3.2. SMOTE\u003c/h2\u003e \u003cp\u003eSMOTE is an oversampling technique that creates synthetic samples for the minority class by interpolating between existing minority samples. SMOTE enables researchers to use synthetic elements to rebalance under-sampled data and is one of the most effective techniques to address imbalanced datasets (Almhaithawi et al., \u003cspan citationid=\"CR7\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e). As shown in Fig.\u0026nbsp;\u003cspan refid=\"Fig3\" class=\"InternalRef\"\u003e3\u003c/span\u003e, SMOTE used the k-nearest neighbor closest to the data points to create synthetic samples from the 2000 fraud instances (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e, p. 18). Instead of randomly oversampling the data with replacements, SMOTE takes \"each minority class sample and introduces synthetic examples\u0026hellip; joining any/all of the k minority class nearest neighbors\" (Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e, p. 327). Depending on the number of instances needed to balance the data, SMOTE will randomly choose to generate synthetic data points from the minority sample by connecting them with their nearest neighbors (Srinilta \u0026amp; Kanharattanachai, \u003cspan citationid=\"CR88\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; B. Sun \u0026amp; Chen, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eSince its publication in 2002, SMOTE has proven successful in a variety of applications across several different domains (Fernandez et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). SMOTE has inspired several approaches to address class imbalance and has significantly contributed to new supervised learning paradigms, including multilabel classification, incremental learning, semi-supervised learning, and multi-instance learning, among others (p. 863). SMOTE has been effectively used to address class imbalance using machine learning algorithms in the financial crime domain with excellent results (M. E. Lokanan, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; M. E. Lokanan \u0026amp; Sharma, \u003cspan citationid=\"CR62\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Others have used SMOTE to address class imbalances within high-dimensional datasets (Maldonado et al., \u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; S. Tiwari et al., \u003cspan citationid=\"CR97\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Maldonado et al. (\u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e2019\u003c/span\u003e) propose a modified version of SMOTE designed for high-dimensional binary scenarios, such as natural language processing. This change involves using a new distance metric that focuses solely on the most important features to generate synthetic observations. Similarly, Tiwari et al. (\u003cspan citationid=\"CR96\" class=\"CitationRef\"\u003e2015\u003c/span\u003e) apply SMOTE to investigate the effect of various resampling ratios on observed peptides and absent peptides in protein mass spectrometry data. Both studies found that class balance greatly improves the performance of machine learning models (Maldonado et al., \u003cspan citationid=\"CR65\" class=\"CitationRef\"\u003e2019\u003c/span\u003e; A. K. Tiwari et al., \u003cspan citationid=\"CR96\" class=\"CitationRef\"\u003e2015\u003c/span\u003e). Others studied how close the distribution of the patterns and number of neighbors generated by SMOTE is to the original distribution and found that it performed better on large rather than small datasets (Elreedy \u0026amp; Atiya, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2019\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec11\" class=\"Section2\"\u003e \u003ch2\u003e3.3. SVMSMOTE\u003c/h2\u003e \u003cp\u003eSupport Vector Machine The Synthetic Minority OverSampling Technique (SVMSMOTE) is a type of oversampling method made for datasets that are not balanced. Its goal is to fix the overfitting problems that come with regular SMOTE (AlJame et al., \u003cspan citationid=\"CR5\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, pp. 4\u0026ndash;5). The SMOTE algorithm and Support Vector Machines (SVM) are combined in SVMSMOTE to make fake samples that are less similar to the original minority class instances (Krayem et al., \u003cspan citationid=\"CR55\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Nguyen et al., \u003cspan citationid=\"CR74\" class=\"CitationRef\"\u003e2011\u003c/span\u003e). To use this method, the original minority class samples are used to train an SVM classifier and find the decision boundary. This is then used to find \"safe\" and \"borderline\" areas within the minority class distribution. Synthetic samples are then subsequently generated by interpolating between instances identified as \"safe,\" thus reducing the risk of overfitting by ensuring the synthetic samples are less similar to the original minority class instances.\u003c/p\u003e \u003cp\u003eSVMSMOTE has been widely used across various domains to address class imbalances with enhanced performance. Studies have utilized SVMSMOTE to address class imbalances in datasets pertaining to medicine, education, and cancer research. Researchers have used SVMSMOTE to conduct diagnostic tests and predict prostate cancer using multiparametric data with outstanding results (Barlow et al., 2019; Bertelli et al., \u003cspan citationid=\"CR20\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Sujitha and Paramasivan (\u003cspan citationid=\"CR91\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) employed SVMSMOTE to predict stages in lung diseases with enhanced performance (Sujitha \u0026amp; Paramasivan, \u003cspan citationid=\"CR91\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). Others have used SVMSMOTE to guess how well students will do in multiple classes using a dataset for education that has better classification performance (Ghorbani \u0026amp; Ghousi, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Tariq et al., \u003cspan citationid=\"CR95\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). All of these studies show that SVMSMOTE can be used to make models work better and fix uneven data distributions in datasets from various fields.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec12\" class=\"Section2\"\u003e \u003ch2\u003e3.4 SMOTE Tomek\u003c/h2\u003e \u003cp\u003eSMOTETomek is a fusion of SMOTE and Tomek links, aimed at balancing imbalanced datasets while reducing noise. Applying the SMOTE algorithm to oversample the minority class is the first step in the process and removing Tomek links refines the dataset (Z. Wang et al., \u003cspan citationid=\"CR100\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). This process aims to eliminate overlapping and noisy observations, thereby enhancing the quality of the dataset and the performance of the machine learning models. By combining the strengths of SMOTE in generating synthetic samples and Tomek links in cleaning the dataset, SMOTETomek creates a more balanced and refined dataset for machine learning tasks.\u003c/p\u003e \u003cp\u003eSMOTETomek has been widely utilized in various domains to address imbalanced datasets and enhance model performance. In healthcare, SMOTETomek has been applied to build models for disease prediction, such as diabetes and hypertension, with improved accuracy and sensitivity (Fitriyani et al., \u003cspan citationid=\"CR35\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Cancer research studies have also used SMOTETomek to address skewed data and accurately predict high-risk prostate and cervical cancer (Boratto et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Ijaz et al., \u003cspan citationid=\"CR48\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Lin et al., \u003cspan citationid=\"CR59\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Tanimu et al., \u003cspan citationid=\"CR94\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). In computer science, SMOTETomek has been used for recommender systems and predicting software bugs, showing that it can be used in a variety of situations and is good at fixing class imbalances (Arif et al., \u003cspan citationid=\"CR11\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Boratto et al., \u003cspan citationid=\"CR21\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). SMOTETomek has also been used effectively to address severe sample distribution imbalances in personality recognition datasets (Z. Wang et al., \u003cspan citationid=\"CR100\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). These studies collectively underscore the utility of SMOTETomek in various domains to address class imbalance issues across various datasets.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec13\" class=\"Section2\"\u003e \u003ch2\u003e3.5 K-MeansSMOTE\u003c/h2\u003e \u003cp\u003eK-MeansSMOTE is an oversampling technique designed to address class imbalance in datasets by generating synthetic samples for the minority class using k-means clustering. This method integrates the k-means clustering algorithm with SMOTE, executed in three distinct steps: clustering, filtering, and oversampling (Chen \u0026amp; Zhang, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Douzas et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). Initially, the algorithm employs k-means clustering to partition the minority class samples into clusters. Subsequently, synthetic samples are created for each cluster by interpolating feature values from the minority class samples within the cluster. These synthetic instances are then added to the original dataset, resulting in a rebalanced dataset (De \u0026amp; Prabu, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eK-MeansSMOTE has proven to be effective in oversampling methods and has improved model performance across various classification datasets (Chen \u0026amp; Zhang, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Xu et al., \u003cspan citationid=\"CR103\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). In medical science, K-MeansSMOTE has been used to effectively balance data. In a recent study, the authors applied K-MeansSMOTE to eight UCI medical datasets with excellent classification scores (Xu et al., \u003cspan citationid=\"CR103\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). In other studies, such as predicting credit default, K-MeansSMOTE has been proven to be very efficient in addressing class imbalance with enhanced performance (Alam et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Chen \u0026amp; Zhang, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Srinivasan et al., \u003cspan citationid=\"CR89\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; T. Sun \u0026amp; Vasarhelyi, \u003cspan citationid=\"CR93\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). K-MeansSMOTE has also been used to handle imbalances in classifying financial distress companies and shows improved performance across various classification metrics (Aljawazneh et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). Another domain where K-MeansSMOTE has been effectively used to address class imbalance and enhance performance is churn prediction (De \u0026amp; Prabu, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec14\" class=\"Section2\"\u003e \u003ch2\u003e3.6 SMOTE\u0026thinsp;+\u0026thinsp;ENN\u003c/h2\u003e \u003cp\u003eSMOTe\u0026thinsp;+\u0026thinsp;ENN is an oversampling technique devised to rectify imbalanced datasets by combining SMOTE with the edited nearest neighbor (ENN) method. The SMOTE\u0026thinsp;+\u0026thinsp;ENN technique operates in two sequential steps. Initially, SMOTE is applied to produce synthetic samples for the minority class. SMOTE accomplishes this by selecting a minority class sample along with its k-nearest neighbors (k-NN), then creating new synthetic samples by interpolating the feature values between the chosen sample and its neighbors. Subsequently, the ENN method is utilized to eliminate any majority-class samples that are misclassified by a k-NN classifier. ENN detects the majority of class samples misclassified by the k-NN classifier and removes them from the dataset. By amalgamating these two steps, SMOTE\u0026thinsp;+\u0026thinsp;ENN can simultaneously generate new synthetic samples for the minority class and eliminate misclassified majority class samples (M. E. Lokanan, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Sisodia et al., \u003cspan citationid=\"CR87\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; F. Yang et al., \u003cspan citationid=\"CR105\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eSMOTE\u0026thinsp;+\u0026thinsp;ENN has proven to be effective in balancing the class distribution and enhancing classification performance. Studies have found that SMOTE\u0026thinsp;+\u0026thinsp;ENN consistently yields superior outcomes compared to most oversampling methods (Batista et al., \u003cspan citationid=\"CR18\" class=\"CitationRef\"\u003e2004\u003c/span\u003e; Singh et al., \u003cspan citationid=\"CR86\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Sisodia et al., \u003cspan citationid=\"CR87\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). SMOTE\u0026thinsp;+\u0026thinsp;ENN has been extensively used to address class imbalances in medical data classification (Lamari et al., \u003cspan citationid=\"CR58\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). In medicine, SMOTE\u0026thinsp;+\u0026thinsp;ENN has been heavily utilized in healthcare for early detection tasks, such as predicting septic shock onset and diagnosing missed abortion, with enhanced diagnostic accuracy (Xu et al., \u003cspan citationid=\"CR102\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; F. Yang et al., \u003cspan citationid=\"CR105\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Others have used SMOTE\u0026thinsp;+\u0026thinsp;ENN in predicting Parkinson's disease and chronic heart failure with very good classification results (Keller \u0026amp; Pandey, \u003cspan citationid=\"CR52\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; K. Wang et al., \u003cspan citationid=\"CR99\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). In the financial crimes\u0026rsquo; arena, SMOTE\u0026thinsp;+\u0026thinsp;ENN has been effectively used to predict fraud with enhanced performance (M. E. Lokanan, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Mienye \u0026amp; Sun, \u003cspan citationid=\"CR66\" class=\"CitationRef\"\u003e2023\u003c/span\u003e). These studies found that SMOTE\u0026thinsp;+\u0026thinsp;ENN was excellent at balancing the datasets and enhancing the robustness of predictive models.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec15\" class=\"Section2\"\u003e \u003ch2\u003e3.7 ADASYN\u003c/h2\u003e \u003cp\u003eADASYN, short for Adaptive Synthetic Sampling, is an oversampling technique specifically designed to address class imbalances in datasets. The algorithm works by generating synthetic samples for the minority class by adaptively adjusting the density of synthetic samples based on the difficulty of the classification problem (Fernandez et al., \u003cspan citationid=\"CR34\" class=\"CitationRef\"\u003e2018\u003c/span\u003e, p. 870). The basic idea behind ADASYN is to generate more synthetic samples in regions where the decision boundary of the minority class is more complex to increase the diversity of the minority class samples (Haibo He et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2008\u003c/span\u003e). This adaptiveness makes ADASYN particularly effective in scenarios where the imbalance between classes is substantial and the classification problem is challenging.\u003c/p\u003e \u003cp\u003eADASYN oversampling has been widely applied across various domains, showcasing its effectiveness in addressing class imbalances and improving model performance. In the field of medicine, researchers have used ADASYN to up-sample features to mitigate missing value concerns in cervical cancer detection and breast cancer diagnosis with exceptional accuracy (Khan et al., \u003cspan citationid=\"CR53\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Kurniawati et al., \u003cspan citationid=\"CR57\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Munshi, \u003cspan citationid=\"CR70\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Others have employed ADASYN in fraud prediction. In one study, the authors found that ADASYN proved to be more advantageous than the traditional SMOTE algorithm in telecom fraud identification (Lu et al., \u003cspan citationid=\"CR64\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). In other studies, ADASYN was used to predict insurance and credit card fraud with enhanced effectiveness on a balanced data set over an unbalanced one (Cinaroglu, \u003cspan citationid=\"CR28\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Muranda et al., \u003cspan citationid=\"CR71\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Singh et al., \u003cspan citationid=\"CR86\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Subudhi \u0026amp; Panigrahi, \u003cspan citationid=\"CR90\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). ADASYN has also proven to be effective in customer churn prediction (Rao et al., \u003cspan citationid=\"CR81\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). These applications underscore the versatility and efficiency of using ADASYN to up-sample imbalanced datasets in different domains.\u003c/p\u003e \u003cp\u003eThe foregoing review indicates that various oversampling techniques are employed to address the challenge of imbalanced datasets in machine learning. Depending on the dataset and domain, different oversampling methods may be employed. RUS aims to balance the class distribution by manipulating minority class observations across different applications with enhanced performance. The SMOTE-based methods used SMOTE as a foundation to generate synthetic samples for the minority class and combined it with additional methods to balance the classes and enhance performance. ADASYN stands out as a unique category by adjusting synthetic sample density based on the complexity of the decision boundary. In doing so, ADASYN addresses class imbalance by focusing on domains where classification appears to be more challenging. Each oversampling technique reviewed provides distinct advantages and can be chosen according to the specific attributes and characteristics of the dataset and the classification problem being addressed.\u003c/p\u003e \u003c/div\u003e"},{"header":"4 Addressing the Impact of Under and OverSampling","content":"\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e \u003ch2\u003e4.1 Negative and Positive Impact of Undersampling\u003c/h2\u003e \u003cp\u003eWhile resampling is a useful technique to address class imbalances, there are some negative impacts associated with both undersampling and oversampling. One of the problems associated with undersampling is that it discards potentially valuable information from the dataset (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e, p. 16), an approach that tends to underfit the data (Oladunni et al., \u003cspan citationid=\"CR76\" class=\"CitationRef\"\u003e2021\u003c/span\u003e, p. 3). Think of a situation where the ratio of class imbalance is 20,000:100, 40,000:100, or 50,000:100. Removing instances to match the minority and majority classes will lead to a significant loss of data. There is no way to preserve the rich information that will be randomly removed from the majority class with undersampling (Jo \u0026amp; Japkowicz, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2004\u003c/span\u003e). The loss of data can make the decision boundary between the minority and majority classes harder to learn from and increase the loss function (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Moreo et al., \u003cspan citationid=\"CR68\" class=\"CitationRef\"\u003e2016\u003c/span\u003e). Another important point is that the sample chosen from the majority class may be systematically biased and not representative of the population (Chakraborty et al., \u003cspan citationid=\"CR23\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). The loss of information results in poor and inaccurate classification performance on the test set.\u003c/p\u003e \u003cp\u003eDespite these drawbacks, undersampling is a useful strategy to address class imbalance problems. The main advantage of undersampling is that it is a valuable approach to reducing the risk of the model being trained and the analysis conducted by the classifier only on the majority class (Yap et al., \u003cspan citationid=\"CR107\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). Machine learning modeling performed only on the majority class will lead to distortion in the performance accuracy (i.e., the model is only reading the majority class) (Jo \u0026amp; Japkowicz, \u003cspan citationid=\"CR50\" class=\"CitationRef\"\u003e2004\u003c/span\u003e; Rao et al., \u003cspan citationid=\"CR81\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). The minority events are harder for machine learning algorithms to predict because there are only a few instances. There is a high probability, for example, that a dataset characterized by a non-fraud to fraud ratio of 10,000:100 will lead to a model trained on the majority class (non-fraud) because there are fewer instances for the algorithm to learn from the data. Undersampling tackles this issue by reducing the number of instances in the majority class to align with the minority class. With large datasets, the use of undersampling helps to reduce computational costs and improve run time (Fujiwara et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e, p. 3). Since undersampling shrinks the data, less training time is needed (p. 3). Consequently, businesses will require less storage space and time to obtain insights from analyzing the data (Zhu et al., \u003cspan citationid=\"CR112\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). To avoid scaling issues associated with excessive data, it is best to discard some to save on computational time and resource usage.\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec18\" class=\"Section2\"\u003e \u003ch2\u003e4.2 Negative and Positive Impact of Oversampling\u003c/h2\u003e \u003cp\u003eSince minority class instances are randomly selected to produce exact copies of the majority class, random oversampling increases the likelihood of overfitting, especially for higher oversampling rates (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e; Fujiwara et al., \u003cspan citationid=\"CR36\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Overfitting occurs when the model achieves high accuracy on the training data (overfits) but is a poor predictor (or fails to perform well) on the unseen test data (Santos et al., \u003cspan citationid=\"CR83\" class=\"CitationRef\"\u003e2018\u003c/span\u003e). The use of artificial data to construct rules for the replication of the minority class that are apparently accurate may end up generating the wrong synthetic instances and increasing the generalization error (Zhu et al., \u003cspan citationid=\"CR112\" class=\"CitationRef\"\u003e2017\u003c/span\u003e, p. 330). A more fundamental problem with oversampling is that it overgeneralizes the minority class with data from the majority class and leads to \"overestimation of minority data\" (Pan et al., \u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e2020\u003c/span\u003e, p. 1215). This strategy is particularly problematic in the case of \"highly skewed class distributions where the minority class examples are very sparse, thus resulting in a greater chance of class mixture\" (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e, p. 19). Considering that the model is evaluating the same samples multiple times, datasets that are highly skewed may result in increased computational costs (Zhu et al., \u003cspan citationid=\"CR112\" class=\"CitationRef\"\u003e2017\u003c/span\u003e, p. 333).\u003c/p\u003e \u003cp\u003eDespite these drawbacks, oversampling can have a positive impact on the results (Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e; Elreedy \u0026amp; Atiya, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The reason for oversampling the data is to ensure that there are enough instances to reduce class bias in the performance evaluation (Naseriparsa \u0026amp; Kashani, \u003cspan citationid=\"CR72\" class=\"CitationRef\"\u003e2014\u003c/span\u003e). A dataset that is trained on the majority class label will result in the classifier only reading the instances from the majority class in the test set. SMOTE-based techniques can directly address this problem because they do not simply oversample the data with duplicates; rather, SMOTE generates synthetic data that is slightly different from the original data (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e). The main idea of oversampling is to improve the classification performance of the minority class and reduce the information loss of the majority class (Gong \u0026amp; Gu, \u003cspan citationid=\"CR38\" class=\"CitationRef\"\u003e2016\u003c/span\u003e, p. 3). In this regard, oversampling effectively helps to provide insights on the minority class samples, which would have been overlooked otherwise. In designing studies with imbalanced data, the minority and majority classes have to be dealt with care to ensure that each class has an equal chance to be analyzed by the classifier. Oversampling improves this outcome by giving instances from the original minority class a higher chance of being equally distributed in the data (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Pan et al., \u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec19\" class=\"Section2\"\u003e \u003ch2\u003e4.3 Deep Adaptive Resampling Technique (DART)\u003c/h2\u003e \u003cp\u003eTo build on and address the limitations of commonly used oversampling techniques in handling class imbalance, this study proposes a novel method: DART. DART differentiates itself from existing resampling strategies\u0026mdash;such as SMOTE, ADASYN, and their hybrid variants\u0026mdash;by integrating two key innovations: (1) the generative power of deep learning through Variational Autoencoders (VAEs) and (2) adaptive sampling via pseudo-labeling of ambiguous instances.\u003c/p\u003e \u003cp\u003eTraditional methods like SMOTE (Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e) and ADASYN (Haibo He et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2008\u003c/span\u003e) use interpolation-based strategies that assume linear relationships in the minority class feature space. While effective in generating additional samples, these techniques often produce synthetic instances that lack diversity or fail to capture the complex, nonlinear structure of real-world data (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Elreedy \u0026amp; Atiya, \u003cspan citationid=\"CR33\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). Furthermore, they do not account for class overlap or decision boundary uncertainty\u0026mdash;issues that frequently arise in high-dimensional, noisy domains. Hybrid approaches like SMOTE-ENN or SMOTE-Tomek attempt to filter noisy samples but still rely on static resampling heuristics, limiting their adaptability to dynamic learning conditions.\u003c/p\u003e \u003cp\u003eDART addresses these limitations through a two-stage architecture:\u003c/p\u003e \u003cp\u003e \u003col\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eDeep Generative Modelling with VAEs\u003c/b\u003e: Unlike traditional methods that rely on local interpolation (e.g., SMOTE) or density-based sampling (e.g., ADASYN), DART leverages VAEs to learn a rich latent representation of the minority class. By capturing the underlying structure of the minority class in a lower-dimensional space, VAEs enable the generation of synthetic samples that are both realistic and diverse. These samples reflect the global characteristics of the data rather than being confined to local neighborhoods. As a result, DART produces more informative and representative examples\u0026mdash;rather than mere duplications or linear interpolations\u0026mdash;which enhance robustness in the presence of noise and nonlinearity.\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003cspan\u003e \u003cli\u003e \u003cp\u003e \u003cb\u003eAdaptive Pseudo-Labeling\u003c/b\u003e: DART incorporates a pseudo-labeling strategy using a Random Forest classifier to identify ambiguous majority class instances near the decision boundary (with predicted probabilities in a confidence band, e.g., 0.45\u0026ndash;0.55). These borderline instances are selectively reassigned to the minority class based on model confidence, thereby improving the class boundary definition, and enhancing generalization (An \u0026amp; Suh, \u003cspan citationid=\"CR9\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003c/li\u003e \u003c/span\u003e \u003c/ol\u003e \u003c/p\u003e \u003cp\u003eBy combining these components, DART moves beyond static, rule-based sampling approaches and introduces a dynamic, data-aware framework that responds to the structure and uncertainty within the data itself. Figure\u0026nbsp;\u003cspan refid=\"Fig4\" class=\"InternalRef\"\u003e1\u003c/span\u003e illustrates this dual mechanism: synthetic samples (blue triangles) generated via VAEs expand the minority class, while pseudo-labeled samples (orange squares) refine decision boundaries, all while preserving the distributional integrity of the majority class (light gray).\u003c/p\u003e \u003cp\u003e \u003c/p\u003e \u003cp\u003eThe existing literature on class imbalance reveals a notable paucity of studies that have integrated deep generative modeling with adaptive sampling techniques. Although previous research has investigated VAE-based data augmentation (e.g., Liu et al., \u003cspan citationid=\"CR60\" class=\"CitationRef\"\u003e2023\u003c/span\u003e) and employed classifiers for the detection of borderline instances (e.g., Ghorbani \u0026amp; Ghousi, \u003cspan citationid=\"CR37\" class=\"CitationRef\"\u003e2020\u003c/span\u003e), these methodologies are typically implemented in isolation. In contrast, DART represents a distinctive synthesis of these strategies, providing a cohesive framework that not only enhances minority representation but also improves classification accuracy in proximity to the decision boundary. Consequently, DART emerges as a novel, modular, and scalable solution that can be seamlessly incorporated into any supervised learning pipeline. DART is particularly advantageous in high-stakes domains such as finance, healthcare, and fraud detection, where imbalanced data and classification inaccuracies can have profound implications. Additionally, the adaptive architecture of DART positions it as a significant advancement in the domain of imbalanced learning.\u003c/p\u003e \u003c/div\u003e"},{"header":"5 Research Design","content":"\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\n \u003cp\u003e\u003cstrong\u003e5.1 \u0026nbsp; \u0026nbsp; \u0026nbsp; Data Collection\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eThis study draws on two datasets focused on credit and loan default prediction. The \u003cstrong\u003ecredit default dataset\u003c/strong\u003e, sourced from the UCI Machine Learning Repository, contains detailed information on credit card clients, including their demographic characteristics, credit usage behaviour, repayment history, and default payment status.\u003ca href=\"#_ftn1\" name=\"_ftnref1\" title=\"\"\u003e\u003c/a\u003e1 The second dataset, used for \u003cstrong\u003eloan default prediction\u003c/strong\u003e, was obtained from Kaggle and provides a complementary perspective by capturing borrower-level financial and transactional information relevant to loan performance. Together, these datasets enable a comprehensive evaluation of resampling techniques across distinct but related domains of consumer credit risk.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003e5.1.2 \u0026nbsp; \u0026nbsp;Dataset 1: Credit Card Default\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eThe credit card dataset includes information about credit card holders, such as their demographic characteristics, credit card usage, repayment history, and default payment status. The dataset consists of the following attributes:\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDemographic Characteristics:\u003c/strong\u003e Demographic attributes gender, age, level of education, and marital status.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCredit Card Usage:\u003c/strong\u003e The variable LIMIT_BAL represent the amount of credit given to each cardholder.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eRepayment History\u003c/strong\u003e: The dataset provides information on the repayment status of cardholders for six consecutive months (PAY_0 to PAY_6). These variables indicate whether payments were made duly or delayed, and if delayed, the duration of the delay.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eBill Statements:\u003c/strong\u003e Amount of bill statements are in reverse order from September to April (BILL_AMT1 to BILL_AMT6).\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003ePrevious Payments\u003c/strong\u003e: \u0026nbsp;Includes the amount of previous payment (PAY_AMT1 to PAY_AMT6) for each cardholder over six months.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eDefault Payment Status\u003c/strong\u003e: The dependent variable default payment indicates whether a cardholder defaulted on their payment in the following month (1=yes, 0=no).\u003c/p\u003e\n\u003cp\u003eThe credit card dataset is commonly used for predictive modelling purposes, specifically to estimate the probability of default based on demographic and repayment history characteristics. Understanding credit risk and devising methods to minimize default rates is crucial for credit card issuers. There are 23 independent variables and data on 30,000 clients in the dataset. Table 1 displays the compilation of independent variables that make up the dataset.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eTable 1:\u003c/strong\u003e Variables and Measurements of Credit Default dataset\u003c/p\u003e\n \u003ctable border=\"0\" cellspacing=\"0\" cellpadding=\"0\" width=\"677\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"bottom\" style=\"width: 15.7817%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eVariable\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 72.5664%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eDescription\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"bottom\" style=\"width: 11.6519%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eType\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.7817%;\"\u003e\n \u003cp\u003eLIMIT_BAL\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72.5664%;\"\u003e\n \u003cp\u003eAmount of given credit in NT dollars (includes individual and family/supplementary credit)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.6519%;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.7817%;\"\u003e\n \u003cp\u003eSEX\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72.5664%;\"\u003e\n \u003cp\u003eGender\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.6519%;\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.7817%;\"\u003e\n \u003cp\u003eEDUCATION\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72.5664%;\"\u003e\n \u003cp\u003eLevel of education (1=graduate school, 2=university, 3=high school, 4=others, 5=unknown, 6=unknown)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.6519%;\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.7817%;\"\u003e\n \u003cp\u003eMARRIAGE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72.5664%;\"\u003e\n \u003cp\u003eMarital status (1=married, 2=single, 3=others)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.6519%;\"\u003e\n \u003cp\u003eNominal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.7817%;\"\u003e\n \u003cp\u003eAGE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72.5664%;\"\u003e\n \u003cp\u003eAge in years\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.6519%;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.7817%;\"\u003e\n \u003cp\u003ePAY_0 to PAY_6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72.5664%;\"\u003e\n \u003cp\u003eRepayment status from April to September 2005 (-1=pay duly, 1=payment delay for one month, 2=payment delay for two months, \u0026hellip; 8=payment delay for eight months, 9=payment delay for nine months and above)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.6519%;\"\u003e\n \u003cp\u003eOrdinal\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.7817%;\"\u003e\n \u003cp\u003eBILL_AMT1 to BILL_AMT6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72.5664%;\"\u003e\n \u003cp\u003eAmount of bill statement from April to September 2005 in NT dollars\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.6519%;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd style=\"width: 15.7817%;\"\u003e\n \u003cp\u003ePAY_AMT1 to PAY_AMT6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 72.5664%;\"\u003e\n \u003cp\u003eAmount of previous payment from April to September 2005 in NT dollars\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd style=\"width: 11.6519%;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003cdiv id=\"Sec23\" class=\"Section3\"\u003e\n \u003ch2\u003e5.1.3 Dataset 2: Load Default\u003c/h2\u003e\n \u003cp\u003eThe loan default dataset used in this study captures a broad range of borrower and loan-related characteristics. The dataset comprises 40 features and 855,968 observations. As shown in Table \u003cspan class=\"InternalRef\"\u003e2\u003c/span\u003e, the data can be grouped into five major categories: (1) Loan and credit information, including loan amount, interest rate, term, and purpose; (2) Borrower demographics and employment details, such as income, homeownership, and employment length; (3) Loan timeline and historical activities, which tracks issuance dates and the borrower\u0026rsquo;s earliest credit line; (4) Credit history and utilization metrics, including the debt-to-income ratio, revolving balances, and account delinquencies; and (5) Loan payment, recovery, and policy features, covering payment behaviour, recoveries, late fees, and compliance with lending policies.\u003c/p\u003e\n \u003cp\u003eGiven the dataset\u0026rsquo;s large size (over 855,000 records), processing the full set required substantial computational resources and time. To address these issues, stratified sampling was employed to reduce the dataset to 30,000 observations while preserving the original distribution of the target variable (loan default vs. non-default). The stratified sample distribution ensured balanced representation of both majority and minority classes, making the data more manageable for training and evaluating machine learning models.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab2\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eVariables and Measurements of Loan Default Dataset\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eVariable(s)\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eType\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eloan_amnt, funded_amnt, funded_amnt_inv\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLoan amount applied for, amount funded, and investor-funded amount\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eterm, int_rate, installment\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLoan term (months), interest rate, and monthly installment payment\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003egrade, sub_grade, purpose\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLoan grade, sub-grade, and purpose of loan (e.g., car, education)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eemp_length, home_ownership, annual_inc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eEmployment duration, homeownership status, and annual income\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMixed (Nominal/Continuous)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003everification_status, application_type\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIncome verification status and application type (individual or joint)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eissue_d, earliest_cr_line, last_pymnt_amnt\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLoan issuance date, earliest credit line date, and last payment amount\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDate / Continuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003edti, delinq_2yrs, inq_last_6mths, open_acc, pub_rec, revol_bal, revol_util, total_acc\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDebt-to-income ratio, credit delinquencies, inquiries, open accounts, public records, revolving balance and utilization, total accounts\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMixed (Continuous/Ordinal)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eout_prncp, out_prncp_inv, total_pymnt, total_pymnt_inv, total_rec_prncp, total_rec_int, total_rec_late_fee\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLoan repayment and recovery details, including outstanding principal and total payments\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003erecoveries, collection_recovery_fee\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRecovery and collection fees after default\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003epymnt_plan, initial_list_status, collections_12_mths_ex_med, policy_code\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003ePayment plan status, listing status, recent collections, and policy code\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003edefault\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eWhether the borrower defaulted on the loan (1\u0026thinsp;=\u0026thinsp;yes, 0\u0026thinsp;=\u0026thinsp;no)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBinary (Target)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec24\" class=\"Section2\"\u003e\n \u003ch2\u003e5.2 Data Cleaning and Processing\u003c/h2\u003e\n \u003cdiv id=\"Sec25\" class=\"Section3\"\u003e\n \u003ch2\u003e5.2.1 Credit Card Dataset\u003c/h2\u003e\n \u003cp\u003eThe dataset required minimal cleaning and pre-processing. We removed certain categories from the \u0026quot;Education\u0026quot; variable, namely \u0026quot;others\u0026quot; and the two \u0026quot;unknowns,\u0026quot; to prevent interpretation issues and redundancy during feature engineering. Additionally, we identified features with a correlation above the 0.70 threshold, indicating potential multicollinearity issues. After several iterations of collinearity checks, we removed the following features from the dataset: BILL_AMT1, BILL_AMT2, BILL_AMT3, BILL_AMT4, BILL_AMT5, BILL_AMT6, PAY_3, PAY_4, PAY_5, PAY_AMT1, PAY_AMT2, PAY_AMT3, PAY_AMT4. To scale the numeric features between 0 and 1, we applied MinMaxScaler. This scaling method was chosen because certain algorithms used in the analysis, such as neural networks and k-NN, require input features to be within this range, and the numerical features did not necessarily follow a normal distribution.\u003c/p\u003e\n \u003c/div\u003e\n \u003cdiv id=\"Sec26\" class=\"Section3\"\u003e\n \u003ch2\u003e5.2.2 Loan Default Dataset\u003c/h2\u003e\n \u003cp\u003eThe loan default dataset underwent extensive cleaning and pre-processing to ensure data quality and modelling reliability. Categorical variables with ordinal significance, such as emp_length, term, and grade, were first encoded numerically to preserve order before imputation. For nominal features like home_ownership, purpose, and verification_status, one-hot encoding was applied, with one dummy variable dropped per feature to avoid multicollinearity. Features with near-zero variance, such as policy_code and acc_now_delinq, were removed due to their limited predictive value. Highly correlated features identified via the correlation matrix and variance inflation factor analysis\u0026mdash;such as funded_amnt, funded_amnt_inv, and total_pymnt_inv\u0026mdash;were dropped to prevent redundancy and collinearity issues. Numeric features were scaled using StandardScaler to standardize ranges across variables, which is essential for gradient-based algorithms and neural networks. Missing values were handled based on data type and context, with ordinal features transformed before imputation to avoid misleading replacements. Finally, a new variable credit_history_length was derived from earliest_cr_line and imputed where necessary to retain temporal credit behaviour information.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec27\" class=\"Section2\"\u003e\n \u003ch2\u003e5.3 Resampling Techniques Applied\u003c/h2\u003e\n \u003cp\u003eTo address class imbalance in the loan and credit default datasets, we applied a variety of commonly used undersampling and oversampling techniques. These resampling strategies were selected based on their prevalence in the literature and their suitability for binary classification tasks with highly skewed class distributions. Table \u003cspan class=\"InternalRef\"\u003e3\u003c/span\u003e summarizes the techniques applied and their core mechanisms.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab3\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eSummary of Resampling Techniques Used\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eTechnique\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eType\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandom Undersampling (RUS)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUndersampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandomly removes majority class samples to achieve class balance\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTomek Links\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUndersampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIdentifies and removes borderline majority samples paired with minority samples\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNearMiss\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUndersampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSelects majority samples that are closest to minority class instances\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCondensed Nearest Neighbor (CNN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUndersampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRetains only majority samples necessary for correct classification\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandom Oversampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOversampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandomly duplicates minority class instances\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSMOTE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOversampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSynthesizes new minority samples by interpolating between nearest neighbors\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSVMSMOTE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOversampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eApplies SVM to generate synthetic samples along decision boundaries\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSMOTE-Tomek\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHybrid\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCombines SMOTE with Tomek Links for oversampling and noise reduction\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eSMOTE-ENN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHybrid\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCombines SMOTE with Edited Nearest Neighbor for enhanced noise filtering\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eK-MeansSMOTE\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOversampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUses clustering before SMOTE to ensure diversity in synthetic samples\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eADASYN\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eOversampling\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGenerates synthetic samples adaptively, focusing on hard-to-learn examples\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eDART (Proposed)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHybrid\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUses a Variational Autoencoder to generate synthetic samples and pseudo-labeling to adaptively include borderline majority instances\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003cdiv id=\"Sec28\" class=\"Section3\"\u003e\n \u003ch2\u003e5.3.1 Over-Sampling with DART\u003c/h2\u003e\n \u003cp\u003eThe DART technique represents an innovative two-stage hybrid oversampling method designed to mitigate class imbalance by integrating deep generative modeling with adaptive pseudo-labeling. In contrast to traditional methods such as SMOTE and ADASYN, which depend on static interpolation or fixed neighborhood heuristics, DART employs a dynamic, data-aware approach to improve the representation of minority classes and refine the delineation of class boundaries.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eStage 1: Generating Synthetic Minority Samples with Variational Autoencoders (VAEs)\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eDART begins by training a VAE on the minority class instances \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{X}_{\\text{min}}\\in\\:{R}^{n\\times\\:d}\\)\u003c/span\u003e\u003c/span\u003e, where \u003cem\u003en\u003c/em\u003e is the number of samples and \u003cem\u003ed\u003c/em\u003e is the number of features. The VAE learns a latent distribution \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:z\\sim\\:\\mathcal{N}\\left(0,I\\right)\\)\u003c/span\u003e\u003c/span\u003e, from which synthetic samples are drawn. These samples are then decoded using:\u003c/p\u003e\n \u003cdiv id=\"Equa\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equa\" name=\"EquationSource\"\u003e$$\\:\\stackrel{\\sim}{x}=\\text{Decoder}\\left(z\\right)$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003cp\u003eThe decoded instances \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\stackrel{\\sim}{{X}_{\\text{syn}}}\\)\u003c/span\u003e\u003c/span\u003e reflect the complex, non-linear structure of the minority class and are used to balance the training data. Unlike traditional oversampling approaches, which often create limited or redundant examples, VAEs generate diverse and structurally realistic samples, improving robustness in the presence of noise and non-linearity.\u003c/p\u003e\n \u003cp\u003e\u003cstrong\u003eStage 2: Adaptive Pseudo-Labeling for Boundary Refinement\u003c/strong\u003e\u003c/p\u003e\n \u003cp\u003eIn the second stage, a Random Forest classifier is trained on the original dataset to estimate the probability that each instance belongs to the minority class:\u003c/p\u003e\n \u003cdiv id=\"Equb\" class=\"Equation\"\u003e\n \u003cdiv class=\"mathdisplay\" id=\"FileID_Equb\" name=\"EquationSource\"\u003e$$\\:\\widehat{p}\\left(x\\right)=P\\left(y=1∣x\\right)$$\u003c/div\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n \u003cp\u003eIn this stage, the majority of class instances with prediction probabilities within the range \u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:0.45\\le\\:\\widehat{p}\\left(x\\right)\\le\\:0.55\\)\u003c/span\u003e\u003c/span\u003e are considered ambiguous. A targeted subset of these borderline cases is pseudo-labeled as minority class and incorporated into the resampled training set. The inclusion of such cases strengthens the model\u0026apos;s ability to distinguish between classes in regions of high uncertainty. The integration of VAE-generated samples with selectively reclassified majority instances enables DART to address both under-representation and boundary ambiguity. Minority class coverage is expanded in a way that maintains data realism, while the decision boundary is simultaneously sharpened through adaptive relabeling. Such a mechanism is particularly valuable in domains characterized by noisy, overlapping, or sparse minority class observations.\u003c/p\u003e\n \u003c/div\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec29\" class=\"Section2\"\u003e\n \u003ch2\u003e5.4 Algorithm Selection\u003c/h2\u003e\n \u003cp\u003eTable \u003cspan class=\"InternalRef\"\u003e4\u003c/span\u003e presents the classification algorithms used to evaluate the performance of various under- and oversampling techniques. These algorithms were deliberately selected to represent a broad spectrum of machine learning categories, including ensemble methods, deep learning, linear models, and probabilistic approaches. Each classifier has a distinct mathematical foundation and learning mechanism, allowing us to examine how different algorithm types respond to class imbalance and resampling strategies. Rather than relying solely on prior literature to guide algorithm selection, we curated a balanced mix to ensure comprehensive assessment of the datasets. Using different classifications algorithms enables a robust comparison of resampling methods across diverse modelling paradigms and works to enhance the reliability and generalizability of our findings.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab4\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 4\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003eAlgorithms Employed\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"4\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eCategory\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eAlgorithm\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eRepresentative Formula\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eEnsemble Learning\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCombines multiple decision trees and uses majority voting for prediction.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\widehat{y}=\\text{mode}\\left({h}_{1}\\left(x\\right),{h}_{2}\\left(x\\right),\\dots\\:,{h}_{T}\\left(x\\right)\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eCatBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eGradient boosting on decision trees with categorical feature handling.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eIterative updates:\u003c/p\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{F}_{t+1}\\left(x\\right)={F}_{t}\\left(x\\right)+{\\eta\\:}\\cdot\\:{h}_{t}\\left(x\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eDeep Learning\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eArtificial Neural Network (ANN)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eUses multiple layers of neurons to learn complex patterns in data.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\widehat{y}={\\sigma\\:}\\left({W}_{2}\\cdot\\:\\text{ReLU}\\left({W}_{1}x+{b}_{1}\\right)+{b}_{2}\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eLinear Models\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eModels log-odds of the target as a linear combination of features.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\widehat{p}=\\frac{1}{1+{e}^{-\\left({{\\beta\\:}}_{0}+{{\\beta\\:}}_{1}{x}_{1}+\\dots\\:+{{\\beta\\:}}_{n}{x}_{n}\\right)}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\u0026nbsp;\u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eStochastic Gradient Descent (SGD)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eMinimizes loss function iteratively using gradient updates.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:{\\theta\\:}={\\theta\\:}-{\\eta\\:}\\cdot\\:{\\nabla\\:}_{{\\theta\\:}}J\\left({\\theta\\:}\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eProbabilistic Model\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eNaive Bayes\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eApplies Bayes\u0026apos; theorem assuming feature independence.\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:P\\left(y∣x\\right)\\propto\\:P\\left(y\\right){\\prod\\:}_{i=1}^{n}P\\left({x}_{i}∣y\\right)\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003c/div\u003e\n\u003cdiv id=\"Sec30\" class=\"Section2\"\u003e\n \u003ch2\u003e5.5 Evaluation Metrics\u003c/h2\u003e\n \u003cp\u003eWe used traditional machine learning classification metrics to evaluate the performance of the models. These include accuracy, precision, recall, F1-score, and area under the ROC curve (AUC)\u0026mdash;all of which are widely recognized and commonly used in the evaluation of classification tasks. In addition, we included the Matthews Correlation Coefficient (MCC), a more balanced metric particularly suitable for evaluating model performance on imbalanced datasets, as it considers all four components of the confusion matrix - true positives (TP), true negatives (TN), false positives (FP), and false negatives (FN). While the details of these metrics are well established in the machine learning literature and not discussed extensively here, the formula for each is presented in Table \u003cspan class=\"InternalRef\"\u003e5\u003c/span\u003e for reference.\u003c/p\u003e\n\u003cdiv class=\"gridtable\"\u003e\n \u003ctable id=\"Tab5\" border=\"1\"\u003e\n \u003ccaption language=\"En\"\u003e\n \u003cdiv class=\"CaptionNumber\"\u003eTable 5\u003c/div\u003e\n \u003cdiv class=\"CaptionContent\"\u003e\n \u003cp\u003ePerformance Metrics\u003c/p\u003e\n \u003c/div\u003e\n \u003c/caption\u003e\n \u003ccolgroup cols=\"3\"\u003e\u003c/colgroup\u003e\n \u003cthead\u003e\n \u003ctr\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eMetric\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eDescription\u003c/p\u003e\n \u003c/th\u003e\n \u003cth align=\"left\"\u003e\n \u003cp\u003eFormula\u003c/p\u003e\n \u003c/th\u003e\n \u003c/tr\u003e\n \u003c/thead\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProportion of correct predictions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTP\u0026thinsp;+\u0026thinsp;TN/TP\u0026thinsp;+\u0026thinsp;TN\u0026thinsp;+\u0026thinsp;FP\u0026thinsp;=\u0026thinsp;FN\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProportion of true positive predictions\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTP/TP\u0026thinsp;+\u0026thinsp;FP\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (Sensitivity)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eProportion of actual positives correctly identified\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eTP/TP\u0026thinsp;+\u0026thinsp;FN\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eF1 Score\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eHarmonic mean of precision and recall\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e2*Precision*Recall/Precision\u0026thinsp;+\u0026thinsp;Recall\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eArea under the ROC Curve\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u0026sum;\u003cem\u003ei\u003c/em\u003e\u0026thinsp;=\u0026thinsp;1\u003cem\u003en\u003c/em\u003e\u0026thinsp;\u0026minus;\u0026thinsp;12(\u003cem\u003exi\u003c/em\u003e\u0026thinsp;+\u0026thinsp;1\u0026minus;\u003cem\u003exi\u003c/em\u003e)\u0026sdot;(\u003cem\u003eyi\u003c/em\u003e\u0026thinsp;+\u0026thinsp;\u003cem\u003eyi\u003c/em\u003e\u0026thinsp;+\u0026thinsp;1)/2\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cstrong\u003eMatthews Correlation Coefficient (MCC)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003eBalanced metric for binary classification\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd align=\"left\"\u003e\n \u003cp\u003e\u003cspan class=\"InlineEquation\"\u003e\u003cspan class=\"mathinline\"\u003e\\(\\:\\frac{TP\\cdot\\:TN-FP\\cdot\\:FN}{\\sqrt{\\left(TP+FP\\right)\\left(TP+FN\\right)\\left(TN+FP\\right)\\left(TN+FN\\right)}}\\)\u003c/span\u003e\u003c/span\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n \u003c/table\u003e\n \u003c/div\u003e\n \u003cp\u003e\u003cbr\u003e\u003c/p\u003e\n\u003c/div\u003e"},{"header":"6 Findings and Analysis","content":"\u003cp\u003eOur objective in looking at these findings is to find out how the under- and oversampling techniques affect the performance of the machine learning algorithms used for data analysis. It is important to note that we are not advocating for a specific method over others. Rather, we believe that the choice of under- or oversampling techniques should be based on the type of data and the desired balance between precision, recall, and overall accuracy, depending on the specific research objectives. Our intention is to present the under- and oversampling techniques that yield the best results in addressing class imbalances using the credit default dataset.\u003c/p\u003e \u003cdiv id=\"Sec32\" class=\"Section2\"\u003e \u003ch2\u003e6.1 Results of Undersampling Techniques\u003c/h2\u003e \u003cdiv id=\"Sec33\" class=\"Section3\"\u003e \u003ch2\u003e6.1.1 Credit Default Dataset\u003c/h2\u003e \u003cp\u003eThe results in Table\u0026nbsp;\u003cspan refid=\"Tab6\" class=\"InternalRef\"\u003e7\u003c/span\u003e provide a comprehensive comparison of oversampling techniques applied to credit default prediction using six machine learning algorithms. Among these methods, K-MeansSMOTE clearly emerged as the most robust and effective, delivering consistently high performance across all classifiers. Under K-MeansSMOTE, classifiers such as ANN, Gradient Descent, Logistic Regression, and CatBoost achieved accuracy scores of 0.88, precision scores above 0.90, and AUC values as high as 0.91 (ANN), with F1-scores all above 0.83. Even Na\u0026iuml;ve Bayes, which typically underperforms in imbalanced contexts, yielded a strong recall of 0.87 with K-MeansSMOTE, demonstrating its enhanced sensitivity to the minority class. These results suggest that K-MeansSMOTE not only balances the dataset effectively but also generates high-quality synthetic samples that preserve class boundaries, leading to better model generalization and discrimination (Chen \u0026amp; Zhang, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; De \u0026amp; Prabu, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Douzas et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2018\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eOther methods showed moderate to inconsistent results. SMOTEENN, for example, performed reasonably well with ANN, producing F1-scores of up to 0.79 and AUC values around 0.87, but struggled with models like CatBoost and Random Forest. ROS and SMOTE achieved mid-range performance across metrics, while ADASYN and SVMSMOTE yielded the weakest outcomes, particularly in precision and AUC, indicating difficulties in generating informative minority samples. Notably, ANN consistently performed well across all resampling methods, further validating its flexibility and adaptability to imbalanced datasets (Chawla et al., \u003cspan citationid=\"CR26\" class=\"CitationRef\"\u003e2002\u003c/span\u003e; Kurani et al., \u003cspan citationid=\"CR56\" class=\"CitationRef\"\u003e2023\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab6\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 7\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eOversampling Sampling Results for Credit Default Dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlgorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1-Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMethod\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.84\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec34\" class=\"Section3\"\u003e \u003ch2\u003e6.1.2 Loan Default Dataset\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab7\" class=\"InternalRef\"\u003e8\u003c/span\u003e displays the oversampling results for the loan default dataset. The evaluation results highlight several important trends across resampling methods and machine learning classifiers. Like the results from the credit default dataset, K-MeansSMOTE consistently outperformed all other resampling techniques, producing the highest F1-Scores (up to 0.71), strong AUC values (up to 0.92), and the highest MCC of 0.73 with both CatBoost and Random Forest classifiers. Notably, CatBoost emerged as the most effective algorithm overall, maintaining top performance across nearly all resampling techniques. For example, under K-MeansSMOTE, SMOTE, and ADASYN, CatBoost achieved the best combination of accuracy, precision, AUC, and MCC. ANN and Gradient Descent classifiers, while generally performing well, showed greater variability depending on the oversampling method used\u0026mdash;highlighting their sensitivity to data distributions. Although SMOTEENN, ROS, and SVMSMOTE provided moderate improvements in minority class recall, their precision and MCC were generally lower than those achieved with K-MeansSMOTE. ADASYN, despite performing decently in recall, struggled to balance precision and F1-Scores across most classifiers. These findings reaffirm that K-MeansSMOTE, when paired with CatBoost or Random Forest, offers the most robust and generalizable solution for handling class imbalance in loan default prediction (Aljawazneh et al., \u003cspan citationid=\"CR6\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; De \u0026amp; Prabu, \u003cspan citationid=\"CR30\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; Srinivasan et al., \u003cspan citationid=\"CR89\" class=\"CitationRef\"\u003e2024\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab7\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 8\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eOversampling Sampling Results for Loan Default Dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eKMeansSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.19\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTEENN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.59\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eROS\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSVMSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.28\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.41\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTETomek\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.26\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eSMOTE\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.2\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.20\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.94\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eADASYN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003cp\u003eFor both the credit and loan default datasets, K-MeansSMOTE emerged as the most effective oversampling technique, consistently delivering superior performance across a wide range of evaluation metrics, including F1-Score, AUC, and MCC. Among the classifiers tested, CatBoost stood out as the top-performing algorithm on both datasets, demonstrating robust precision, strong recall, and excellent generalization capability across different resampling strategies (J. T. Hancock \u0026amp; Khoshgoftaar, \u003cspan citationid=\"CR42\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Huang et al., \u003cspan citationid=\"CR47\" class=\"CitationRef\"\u003e2019\u003c/span\u003e). The combination of K-MeansSMOTE with CatBoost offers a powerful and scalable approach to addressing class imbalance and achieving high predictive performance in credit risk modelling tasks. These findings suggest that practitioners dealing with imbalanced financial datasets should strongly consider the K-MeansSMOTE with CatBoost pairing for reliable and accurate classification outcomes.\u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec35\" class=\"Section2\"\u003e \u003ch2\u003e6.2 Results of Undersampling Techniques\u003c/h2\u003e \u003cdiv id=\"Sec36\" class=\"Section3\"\u003e \u003ch2\u003e6.2.1 Credit Default Dataset\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab8\" class=\"InternalRef\"\u003e9\u003c/span\u003e provides the results for the undersampling techniques of the credit default dataset. The results reveal that NearMiss outperformed other methods in balancing predictive performance across models. Particularly, when combined with ANN, NearMiss achieved the best overall performance, with an F1-Score of 0.81, AUC of 0.87, and strong precision and recall (0.83 and 0.80, respectively). These results suggests that NearMiss is an effective technique to preserve informative patterns from the minority class while reducing the majority class (Ha \u0026amp; Lee, \u003cspan citationid=\"CR39\" class=\"CitationRef\"\u003e2016\u003c/span\u003e; Mqadi et al., \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). TomekLinks, in contrast, exhibited lower recall across all classifiers\u0026mdash;highlighting its limited utility in retaining true positive instances (Ai-jun \u0026amp; Peng, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Devi et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e). For example, while CatBoost under TomekLinks attained a high precision of 0.76, its recall was only 0.31, resulting in a low F1-score of 0.44. Random undersampling yielded moderate performance across classifiers, with F1-scores hovering around 0.62\u0026ndash;0.66, suggesting it may be too aggressive in discarding data. CNN, similarly, underperformed in recall, especially with models like ANN and Gradient Descent, where recall dropped to 0.42 and 0.48, respectively.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab8\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 9\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eUndersampling Sampling Results for Credit Default Dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"7\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlgorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1_Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMethod\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.31\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.44\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.32\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.45\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.61\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.75\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.82\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.8\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.51\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.43\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.54\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.70\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003cdiv id=\"Sec37\" class=\"Section3\"\u003e \u003ch2\u003e6.2.2 Loan Default Dataset\u003c/h2\u003e \u003cp\u003eAs shown in Table\u0026nbsp;\u003cspan refid=\"Tab9\" class=\"InternalRef\"\u003e10\u003c/span\u003e, the results from applying various undersampling techniques to the loan default dataset reveal substantial variation in model performance depending on the method used. TomekLinks emerged as the most effective undersampling technique overall, delivering consistently high scores across classifiers such as CatBoost, Random Forest, Logistic Regression, and ANN. These combinations achieved F1-scores up to 0.71 and MCC values as high as 0.73, indicating balanced precision-recall trade-offs and strong alignment between predicted and true labels. Special attention is given to the F1-score and the MCC, which are more appropriate for evaluating imbalanced datasets.\u003c/p\u003e \u003cp\u003eIn contrast, NearMiss performed poorly across all classifiers, with significant drops in F1-scores (as low as 0.13 to 0.25) and MCC values barely exceeding 0.24. Random undersampling produced moderate results, with Na\u0026iuml;ve Bayes achieving the best MCC (0.63) and F1-score (0.64), while most other classifiers struggled. CNN showed stronger performance, especially when paired with Random Forest and CatBoost, achieving F1-scores of 0.71 and MCC values of 0.72 and 0.73 respectively\u0026mdash;comparable to those attained by TomekLinks. Despite CNN\u0026rsquo;s competitive performance, the consistent superiority of TomekLinks across nearly all evaluation metrics and models establishes it as the most robust and reliable undersampling method for the loan default dataset (see Alamri \u0026amp; Ykhlef, \u003cspan citationid=\"CR3\" class=\"CitationRef\"\u003e2024\u003c/span\u003e; Devi et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Pereira et al., \u003cspan citationid=\"CR78\" class=\"CitationRef\"\u003e2020\u003c/span\u003e).\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab9\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 10\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eUndersampling Sampling Results for Loan Default Dataset\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlgorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1_Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMCC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eMethod\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.50\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.60\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.58\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eTomekLinks\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.42\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.14\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.12\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.88\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.10\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.68\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.24\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.49\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.09\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.80\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.47\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.08\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.81\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.15\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.07\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.13\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.11\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eNearMiss\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.37\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.25\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.74\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.91\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.38\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.23\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.35\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.36\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.83\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.21\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.76\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.33\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.89\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.34\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.16\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.79\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.27\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.87\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.29\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eRandom\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.56\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.92\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.72\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.55\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.71\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.73\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.78\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.57\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.66\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.65\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.52\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.67\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.90\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.69\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.77\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.53\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.63\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.85\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.62\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e \u003cp\u003e0.93\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e \u003cp\u003e0.39\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e \u003cp\u003e0.64\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e \u003cp\u003e0.48\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e \u003cp\u003e0.86\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e \u003cp\u003e0.46\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eCNN\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e \u003c/div\u003e \u003cdiv id=\"Sec38\" class=\"Section2\"\u003e \u003ch2\u003e6.3 DART Oversampling\u003c/h2\u003e \u003cp\u003eTable\u0026nbsp;\u003cspan refid=\"Tab10\" class=\"InternalRef\"\u003e11\u003c/span\u003e presents the results of the DART oversampling method applied to both the credit and loan default datasets. As evidenced by the results, DART performed exceptionally well across all classifiers and evaluation metrics. Given the imbalanced nature of the datasets, traditional accuracy alone is not a sufficient indicator of model performance. Therefore, emphasis is placed on more robust evaluation metrics\u0026mdash;F1-Score, AUC, and the MCC \u0026mdash;which are better suited for imbalanced classification problems (Y. Bao et al., \u003cspan citationid=\"CR15\" class=\"CitationRef\"\u003e2022\u003c/span\u003e; M. Lokanan \u0026amp; Sharma, \u003cspan citationid=\"CR63\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). Remarkably, DART achieved scores exceeding 0.98 for all three of these metrics across all classifiers in both datasets. The consistency of these results underscores DART\u0026rsquo;s effectiveness in generating high-quality synthetic samples that support generalization across multiple algorithms. Notably, CatBoost and Random Forest emerged as the top-performing classifiers under the DART algorithm, reinforcing their ability to capitalize on the enhanced balance and diversity introduced by the resampling technique. These findings highlight DART as a state-of-the-art oversampling solution with superior performance and broad applicability in predictive modelling tasks involving class imbalance.\u003c/p\u003e \u003cp\u003e \u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab10\" border=\"1\"\u003e \u003ccaption language=\"En\"\u003e \u003cdiv class=\"CaptionNumber\"\u003eTable 11\u003c/div\u003e \u003cdiv class=\"CaptionContent\"\u003e \u003cp\u003eDART Oversampling Results\u003c/p\u003e \u003c/div\u003e \u003c/caption\u003e \u003ccolgroup cols=\"8\"\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e \u003cdiv align=\"left\" class=\"colspec\" colname=\"c8\" colnum=\"8\"\u003e\u003c/div\u003e \u003cthead\u003e \u003ctr\u003e \u003cth align=\"left\" colspan=\"8\" nameend=\"c8\" namest=\"c1\"\u003e \u003cp\u003eDART Credit Default\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003ctr\u003e \u003cth align=\"left\" colname=\"c1\"\u003e \u003cp\u003eAlgorithm\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c2\"\u003e \u003cp\u003eAccuracy\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c3\"\u003e \u003cp\u003ePrecision\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c4\"\u003e \u003cp\u003eRecall\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c5\"\u003e \u003cp\u003eF1-Score\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c6\"\u003e \u003cp\u003eAUC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c7\"\u003e \u003cp\u003eMCC\u003c/p\u003e \u003c/th\u003e \u003cth align=\"left\" colname=\"c8\"\u003e \u003cp\u003eMethod\u003c/p\u003e \u003c/th\u003e \u003c/tr\u003e \u003c/thead\u003e \u003ctbody\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.100\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colspan=\"8\" nameend=\"c8\" namest=\"c1\"\u003e \u003cp\u003e\u003cb\u003eDART Loan\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003e\u003cb\u003eAlgorithm\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e\u003cb\u003ePrecision\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e\u003cb\u003eRecall\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e\u003cb\u003eF1-Score\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e\u003cb\u003eAUC\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e\u003cb\u003eMCC\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003e\u003cb\u003eMethod\u003c/b\u003e\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eCatBoost\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.998\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eRandom Forest\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.998\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eNa\u0026iuml;ve Bayes\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.95\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.96\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eLogistic Regression\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eGradient Descent\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.97\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003ctr\u003e \u003ctd align=\"left\" colname=\"c1\"\u003e \u003cp\u003eANN\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c2\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c3\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c4\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c5\"\u003e \u003cp\u003e0.99\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c6\"\u003e \u003cp\u003e1.00\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c7\"\u003e \u003cp\u003e0.98\u003c/p\u003e \u003c/td\u003e \u003ctd align=\"left\" colname=\"c8\"\u003e \u003cp\u003eDART\u003c/p\u003e \u003c/td\u003e \u003c/tr\u003e \u003c/tbody\u003e \u003c/colgroup\u003e \u003c/table\u003e\u003c/div\u003e \u003c/p\u003e \u003c/div\u003e"},{"header":"7 Discussion and Conclusion","content":"\u003cp\u003eThe aim of this study was to address the pervasive issue of class imbalance in predictive modeling by evaluating the effectiveness of both advanced undersampling and oversampling techniques. Using two large and highly imbalanced datasets\u0026mdash;one on credit default and the other on loan default\u0026mdash;the study systematically assessed how different resampling strategies impact model performance across a variety of machine learning algorithms, including ensemble methods, probabilistic classifiers, linear models, and deep learning approaches. A central component of this investigation was the introduction of a novel oversampling technique, DART, which was benchmarked against established methods to determine its robustness, adaptability, and generalizability across different datasets and model types.\u003c/p\u003e \u003cp\u003eIn the case of undersampling, NearMiss emerged as the most effective technique for the credit default dataset, delivering balanced F1-scores and AUC values across various classifiers. These result corroborates prior research emphasizing NearMiss\u0026rsquo;s utility in enhancing classifier performance through selective removal of overlapping majority samples (Mqadi et al., \u003cspan citationid=\"CR69\" class=\"CitationRef\"\u003e2021\u003c/span\u003e; Alamsyah et al., \u003cspan citationid=\"CR4\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Meanwhile, TomekLinks outperformed all other undersampling methods in the loan default dataset, showing superior results in terms of F1-score and MCC, especially when paired with CatBoost and Random Forest (Ai-jun \u0026amp; Peng, \u003cspan citationid=\"CR1\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Devi et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Pereira et al., \u003cspan citationid=\"CR78\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). These findings are consistent with the literature highlighting TomekLinks\u0026rsquo; ability to clean noisy boundary points and clarify decision margins (Devi et al., \u003cspan citationid=\"CR31\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Zeng et al., \u003cspan citationid=\"CR108\" class=\"CitationRef\"\u003e2016\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eWhen analyzing oversampling techniques, K-MeansSMOTE consistently delivered the best results across all classifiers in both datasets, based on a range of established evaluation metrics. Oversampling sampling with K-MeansSMOTE achieved high scores in F1, AUC, and MCC\u0026mdash;metrics that are particularly suited for evaluating imbalanced datasets (Douzas et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Chen \u0026amp; Zhang, \u003cspan citationid=\"CR27\" class=\"CitationRef\"\u003e2021\u003c/span\u003e). These findings support the use of K-MeansSMOTE in domains like finance and healthcare, where class boundaries are complex and misclassification is costly (Alam et al., \u003cspan citationid=\"CR2\" class=\"CitationRef\"\u003e2020\u003c/span\u003e; Douzas et al., \u003cspan citationid=\"CR32\" class=\"CitationRef\"\u003e2018\u003c/span\u003e; Srinivasan et al., \u003cspan citationid=\"CR89\" class=\"CitationRef\"\u003e2024\u003c/span\u003e). SMOTE\u0026thinsp;+\u0026thinsp;ENN also performed well in certain contexts, particularly by enhancing precision and reducing class overlap, corroborating previous findings in medical and fraud detection research (M. E. Lokanan, \u003cspan citationid=\"CR61\" class=\"CitationRef\"\u003e2023\u003c/span\u003e; Sisodia et al., \u003cspan citationid=\"CR87\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; F. Yang et al., \u003cspan citationid=\"CR105\" class=\"CitationRef\"\u003e2022\u003c/span\u003e). Conversely, ADASYN tended to underperform in both datasets, reinforcing concerns from prior studies about its sensitivity to noise in highly imbalanced scenarios (Haibo He et al., \u003cspan citationid=\"CR40\" class=\"CitationRef\"\u003e2008\u003c/span\u003e; Singh et al., \u003cspan citationid=\"CR86\" class=\"CitationRef\"\u003e2022\u003c/span\u003e).\u003c/p\u003e \u003cp\u003eMost notably, DART outperformed all traditional oversampling methods across all evaluation metrics. With F1-scores, AUC, and MCC values consistently above 0.97 for both datasets, DART demonstrated an exceptional ability to synthesize informative samples while preserving the underlying distribution of the data. These results underscore the advantage of deep adaptive learning techniques in resampling tasks and open new avenues for improving classification in highly imbalanced domains.\u003c/p\u003e \u003cp\u003eOverall, the results reaffirm the critical role of choosing appropriate resampling techniques based on dataset characteristics, domain constraints, and the performance metric most aligned with the application context (Branco et al., \u003cspan citationid=\"CR22\" class=\"CitationRef\"\u003e2017\u003c/span\u003e; Pan et al., \u003cspan citationid=\"CR77\" class=\"CitationRef\"\u003e2020\u003c/span\u003e). Our study contributes to the growing body of knowledge by empirically validating and benchmarking resampling techniques across diverse classifiers and datasets. Moreover, our study introduces a novel, highly effective oversampling techniques \u0026mdash;DART\u0026mdash;which when tested across two different datasets offers a promising new direction for future work in imbalanced classification.\u003c/p\u003e \u003cdiv id=\"Sec40\" class=\"Section2\"\u003e \u003ch2\u003e7.1 Limitations and Future Research\u003c/h2\u003e \u003cp\u003eDespite its contributions, this study is not without limitations. First, while DART demonstrates promising results, its current implementation is limited to oversampling the minority class. As such, it does not address scenarios where undersampling or hybrid sampling may be more appropriate, particularly in cases where the majority class contains significant noise or redundancy. Limiting DART to oversampling of the minority class limits it applicability to tasks where generating synthetic minority examples is suitable and may not perform optimally in datasets that would benefit from a combination of oversampling and majority class reduction. Future work should explore extending DART to support undersampling strategies or integrating it into more flexible hybrid frameworks that adapt to varying data distributions.\u003c/p\u003e \u003cp\u003eSecond, while this paper employed a range of widely accepted evaluation metrics\u0026mdash;including accuracy, precision, recall, F1-score, AUC, and the MCC\u0026mdash;it did not exhaustively explore the full range of experimental setups and performance measures commonly used in predictive machine learning research. We employed a standard train/test split approach to evaluate model performance; however, alternative validation techniques such as k-fold cross-validation and bootstrapped sample could offer more robust estimates and help mitigate variance due to data partitioning. Future research would benefit from incorporating cross-validation, bootstrapped sampling, or other advanced resampling techniques to enhance the reliability and generalizability of model performance.\u003c/p\u003e \u003c/div\u003e"},{"header":"Declarations","content":"\u003cp\u003e \u003ch2\u003eConflict of interest\u003c/h2\u003e \u003cp\u003eThere is no conflict of interest.\u003c/p\u003e \u003c/p\u003e\u003ch2\u003eFunding statement\u003c/h2\u003e \u003cp\u003eThere is no funding associated with this paper.\u003c/p\u003e\u003ch2\u003eAcknowledgments\u003c/h2\u003e \u003cp\u003eI would like to acknowledge the contribution made by my graduate research assistant, Vikas Maddhesia, to this project.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e \u003cp\u003eData for this paper will be available upon request.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eAi-jun, L., \u0026amp; Peng, Z. (2020). Research on Unbalanced Data Processing Algorithm Base Tomeklinks-Smote. \u003cem\u003eProceedings of the 2020 3rd International Conference on Artificial Intelligence and Pattern Recognition\u003c/em\u003e, 13\u0026ndash;17. https://doi.org/10.1145/3430199.3430222\u003c/li\u003e\n\u003cli\u003eAlam, T. M., Shaukat, K., Hameed, I. A., Luo, S., Sarwar, M. U., Shabbir, S., Li, J., \u0026amp; Khushi, M. (2020). An Investigation of Credit Card Default Prediction in the Imbalanced Datasets. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e, 201173\u0026ndash;201198. https://doi.org/10.1109/ACCESS.2020.3033784\u003c/li\u003e\n\u003cli\u003eAlamri, M., \u0026amp; Ykhlef, M. (2024). Hybrid Undersampling and Oversampling for Handling Imbalanced Credit Card Data. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e12\u003c/em\u003e, 14050\u0026ndash;14060. https://doi.org/10.1109/ACCESS.2024.3357091\u003c/li\u003e\n\u003cli\u003eAlamsyah, A. R. B., Anisa, S. R., Belinda, N. S., \u0026amp; Setiawan, A. (2022). SMOTE and Nearmiss Methods for Disease Classification with Unbalanced Data: Case Study: IFLS 5. \u003cem\u003eProceedings of The International Conference on Data Science and Official Statistics\u003c/em\u003e, \u003cem\u003e2021\u003c/em\u003e(1), 305\u0026ndash;314. https://doi.org/10.34123/icdsos.v2021i1.240\u003c/li\u003e\n\u003cli\u003eAlJame, M., Imtiaz, A., Ahmad, I., \u0026amp; Mohammed, A. (2021). Deep forest model for diagnosing COVID-19 from routine blood tests. \u003cem\u003eScientific Reports\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e(1), 16682. https://doi.org/10.1038/s41598-021-95957-w\u003c/li\u003e\n\u003cli\u003eAljawazneh, H., Mora, A. M., Garcia-Sanchez, P., \u0026amp; Castillo-Valdivieso, P. A. (2021). Comparing the Performance of Deep Learning Methods to Predict Companies\u0026rsquo; Financial Failure. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 97010\u0026ndash;97038. https://doi.org/10.1109/ACCESS.2021.3093461\u003c/li\u003e\n\u003cli\u003eAlmhaithawi, D., Jafar, A., \u0026amp; Aljnidi, M. (2020). Example-dependent cost-sensitive credit cards fraud detection using SMOTE and Bayes minimum risk. \u003cem\u003eSN Applied Sciences\u003c/em\u003e, \u003cem\u003e2\u003c/em\u003e(9), 1574. https://doi.org/10.1007/s42452-020-03375-w\u003c/li\u003e\n\u003cli\u003eAlsowail, R. A. (2022). An Insider Threat Detection Model Using One-Hot Encoding and Near-Miss Under-Sampling Techniques. In M. S. Uddin, P. K. Jamwal, \u0026amp; J. C. Bansal (Eds.), \u003cem\u003eProceedings of International Joint Conference on Advances in Computational Intelligence\u003c/em\u003e (pp. 183\u0026ndash;196). Springer Nature Singapore. https://doi.org/10.1007/978-981-19-0332-8_13\u003c/li\u003e\n\u003cli\u003eAn, B., \u0026amp; Suh, Y. (2020). Identifying financial statement fraud with decision rules obtained from Modified Random Forest. \u003cem\u003eData Technologies and Applications\u003c/em\u003e, \u003cem\u003e54\u003c/em\u003e(2), 235\u0026ndash;255. https://doi.org/10.1108/DTA-11-2019-0208\u003c/li\u003e\n\u003cli\u003eAraf, I., Idri, A., \u0026amp; Chairi, I. (2024). Cost-sensitive learning for imbalanced medical data: A review. \u003cem\u003eArtificial Intelligence Review\u003c/em\u003e, \u003cem\u003e57\u003c/em\u003e(4), 80. https://doi.org/10.1007/s10462-023-10652-8\u003c/li\u003e\n\u003cli\u003eArif, M., Fang, G., Fida, H., Musleh, S., Yu, D.-J., \u0026amp; Alam, T. (2024). iMRSAPred: Improved Prediction of Anti-MRSA Peptides Using Physicochemical and Pairwise Contact-Energy Properties of Amino Acids. \u003cem\u003eACS Omega\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e(2), 2874\u0026ndash;2883. https://doi.org/10.1021/acsomega.3c08303\u003c/li\u003e\n\u003cli\u003eBagui, S., \u0026amp; Li, K. (2021). Resampling imbalanced data for network intrusion detection datasets. \u003cem\u003eJournal of Big Data\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e(1), 6. https://doi.org/10.1186/s40537-020-00390-x\u003c/li\u003e\n\u003cli\u003eBansal, A., \u0026amp; Jain, A. (2021). Analysis of Focussed Under-Sampling Techniques with Machine Learning Classifiers. \u003cem\u003e2021 IEEE/ACIS 19th International Conference on Software Engineering Research, Management and Applications (SERA)\u003c/em\u003e, 91\u0026ndash;96. https://doi.org/10.1109/SERA51205.2021.9509270\u003c/li\u003e\n\u003cli\u003eBao, L., Juan, C., Li, J., \u0026amp; Zhang, Y. (2016). Boosted Near-miss Under-sampling on SVM ensembles for concept detection in large-scale imbalanced datasets. \u003cem\u003eNeurocomputing\u003c/em\u003e, \u003cem\u003e172\u003c/em\u003e, 198\u0026ndash;206. https://doi.org/10.1016/j.neucom.2014.05.096\u003c/li\u003e\n\u003cli\u003eBao, Y., Hilary, G., \u0026amp; Ke, B. (2022). Artificial Intelligence and Fraud Detection. In V. Babich, J. R. Birge, \u0026amp; G. Hilary (Eds.), \u003cem\u003eInnovative Technology at the Interface of Finance and Operations\u003c/em\u003e (Vol. 11, pp. 223\u0026ndash;247). Springer International Publishing. https://doi.org/10.1007/978-3-030-75729-8_8\u003c/li\u003e\n\u003cli\u003eBarlow, Mao, \u0026amp; Khushi. (2019). Predicting High-Risk Prostate Cancer Using Machine Learning Methods. \u003cem\u003eData\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(3), 129. https://doi.org/10.3390/data4030129\u003c/li\u003e\n\u003cli\u003eBasit, M. S., Khan, A., Farooq, O., Khan, Y. U., \u0026amp; Shameem, M. (2022). Handling Imbalanced and Overlapped Medical Datasets: A Comparative Study. \u003cem\u003e2022 5th International Conference on Multimedia, Signal Processing and Communication Technologies (IMPACT)\u003c/em\u003e, 1\u0026ndash;7. https://doi.org/10.1109/IMPACT55510.2022.10029111\u003c/li\u003e\n\u003cli\u003eBatista, G. E. A. P. A., Prati, R. C., \u0026amp; Monard, M. C. (2004). A study of the behavior of several methods for balancing machine learning training data. \u003cem\u003eACM SIGKDD Explorations Newsletter\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e(1), 20\u0026ndash;29. https://doi.org/10.1145/1007730.1007735\u003c/li\u003e\n\u003cli\u003eBauder, R., \u0026amp; Khoshgoftaar, T. (2018). Medicare Fraud Detection Using Random Forest with Class Imbalanced Big Data. \u003cem\u003e2018 IEEE International Conference on Information Reuse and Integration (IRI)\u003c/em\u003e, 80\u0026ndash;87. https://doi.org/10.1109/IRI.2018.00019\u003c/li\u003e\n\u003cli\u003eBertelli, E., Mercatelli, L., Marzi, C., Pachetti, E., Baccini, M., Barucci, A., Colantonio, S., Gherardini, L., Lattavo, L., Pascali, M. A., Agostini, S., \u0026amp; Miele, V. (2022). Machine and Deep Learning Prediction Of Prostate Cancer Aggressiveness Using Multiparametric MRI. \u003cem\u003eFrontiers in Oncology\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e, 802964. https://doi.org/10.3389/fonc.2021.802964\u003c/li\u003e\n\u003cli\u003eBoratto, L., Carta, S., Iguider, W., Mulas, F., \u0026amp; Pilloni, P. (2022). Fair performance-based user recommendation in eCoaching systems. \u003cem\u003eUser Modeling and User-Adapted Interaction\u003c/em\u003e, \u003cem\u003e32\u003c/em\u003e(5), 839\u0026ndash;881. https://doi.org/10.1007/s11257-022-09339-6\u003c/li\u003e\n\u003cli\u003eBranco, P., Torgo, L., \u0026amp; Ribeiro, R. P. (2017). A Survey of Predictive Modeling on Imbalanced Domains. \u003cem\u003eACM Computing Surveys\u003c/em\u003e, \u003cem\u003e49\u003c/em\u003e(2), 1\u0026ndash;50. https://doi.org/10.1145/2907070\u003c/li\u003e\n\u003cli\u003eChakraborty, J., Majumder, S., \u0026amp; Menzies, T. (2021). Bias in machine learning software: Why? how? what to do? \u003cem\u003eProceedings of the 29th ACM Joint Meeting on European Software Engineering Conference and Symposium on the Foundations of Software Engineering\u003c/em\u003e, 429\u0026ndash;440. https://doi.org/10.1145/3468264.3468537\u003c/li\u003e\n\u003cli\u003eChaplot, A., Choudhary, N., \u0026amp; Jain, K. (2019). \u003cem\u003eA Review on Data Level Approaches for Managing Imbalanced Classification Problem\u003c/em\u003e. \u003cem\u003e6\u003c/em\u003e(2), 91\u0026ndash;97.\u003c/li\u003e\n\u003cli\u003eCharte, F., Rivera, A. J., del Jesus, M. J., \u0026amp; Herrera, F. (2015). Addressing imbalance in multilabel classification: Measures and random resampling algorithms. \u003cem\u003eNeurocomputing\u003c/em\u003e, \u003cem\u003e163\u003c/em\u003e, 3\u0026ndash;16. https://doi.org/10.1016/j.neucom.2014.08.091\u003c/li\u003e\n\u003cli\u003eChawla, N. V., Bowyer, K. W., Hall, L. O., \u0026amp; Kegelmeyer, W. P. (2002). SMOTE: Synthetic Minority Over-sampling Technique. \u003cem\u003eJournal of Artificial Intelligence Research\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e, 321\u0026ndash;357. https://doi.org/10.1613/jair.953\u003c/li\u003e\n\u003cli\u003eChen, Y., \u0026amp; Zhang, R. (2021). Research on Credit Card Default Prediction Based on k-Means SMOTE and BP Neural Network. \u003cem\u003eComplexity\u003c/em\u003e, \u003cem\u003e2021\u003c/em\u003e, 1\u0026ndash;13. https://doi.org/10.1155/2021/6618841\u003c/li\u003e\n\u003cli\u003eCinaroglu, S. (2020). Modelling unbalanced catastrophic health expenditure data by using machine‐learning methods. \u003cem\u003eIntelligent Systems in Accounting, Finance and Management\u003c/em\u003e, \u003cem\u003e27\u003c/em\u003e(4), 168\u0026ndash;181. https://doi.org/10.1002/isaf.1483\u003c/li\u003e\n\u003cli\u003eDal Pozzolo, A., Caelen, O., Le Borgne, Y.-A., Waterschoot, S., \u0026amp; Bontempi, G. (2014). Learned lessons in credit card fraud detection from a practitioner perspective. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e41\u003c/em\u003e(10), 4915\u0026ndash;4928. https://doi.org/10.1016/j.eswa.2014.02.026\u003c/li\u003e\n\u003cli\u003eDe, S., \u0026amp; Prabu, P. (2022). A Sampling-Based Stack Framework for Imbalanced Learning in Churn Prediction. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e, 68017\u0026ndash;68028. https://doi.org/10.1109/ACCESS.2022.3185227\u003c/li\u003e\n\u003cli\u003eDevi, D., Biswas, S. kr., \u0026amp; Purkayastha, B. (2017). Redundancy-driven modified Tomek-link based undersampling: A solution to class imbalance. \u003cem\u003ePattern Recognition Letters\u003c/em\u003e, \u003cem\u003e93\u003c/em\u003e, 3\u0026ndash;12. https://doi.org/10.1016/j.patrec.2016.10.006\u003c/li\u003e\n\u003cli\u003eDouzas, G., Bacao, F., \u0026amp; Last, F. (2018). Improving imbalanced learning through a heuristic oversampling method based on k-means and SMOTE. \u003cem\u003eInformation Sciences\u003c/em\u003e, \u003cem\u003e465\u003c/em\u003e, 1\u0026ndash;20. https://doi.org/10.1016/j.ins.2018.06.056\u003c/li\u003e\n\u003cli\u003eElreedy, D., \u0026amp; Atiya, A. F. (2019). A Comprehensive Analysis of Synthetic Minority Oversampling Technique (SMOTE) for handling class imbalance. \u003cem\u003eInformation Sciences\u003c/em\u003e, \u003cem\u003e505\u003c/em\u003e, 32\u0026ndash;64. https://doi.org/10.1016/j.ins.2019.07.070\u003c/li\u003e\n\u003cli\u003eFernandez, A., Garcia, S., Herrera, F., \u0026amp; Chawla, N. V. (2018). SMOTE for Learning from Imbalanced Data: Progress and Challenges, Marking the 15-year Anniversary. \u003cem\u003eJournal of Artificial Intelligence Research\u003c/em\u003e, \u003cem\u003e61\u003c/em\u003e, 863\u0026ndash;905. https://doi.org/10.1613/jair.1.11192\u003c/li\u003e\n\u003cli\u003eFitriyani, N. L., Syafrudin, M., Alfian, G., \u0026amp; Rhee, J. (2019). Development of Disease Prediction Model Based on Ensemble Learning Approach for Diabetes and Hypertension. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e, 144777\u0026ndash;144789. https://doi.org/10.1109/ACCESS.2019.2945129\u003c/li\u003e\n\u003cli\u003eFujiwara, K., Huang, Y., Hori, K., Nishioji, K., Kobayashi, M., Kamaguchi, M., \u0026amp; Kano, M. (2020). Over- and Under-sampling Approach for Extremely Imbalanced and Small Minority Data Problem in Health Record Analysis. \u003cem\u003eFrontiers in Public Health\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e, 178. https://doi.org/10.3389/fpubh.2020.00178\u003c/li\u003e\n\u003cli\u003eGhorbani, R., \u0026amp; Ghousi, R. (2020). Comparing Different Resampling Methods in Predicting Students\u0026rsquo; Performance Using Machine Learning Techniques. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e, 67899\u0026ndash;67911. https://doi.org/10.1109/ACCESS.2020.2986809\u003c/li\u003e\n\u003cli\u003eGong, C., \u0026amp; Gu, L. (2016). A Novel SMOTE-Based Classification Approach to Online Data Imbalance Problem. \u003cem\u003eMathematical Problems in Engineering\u003c/em\u003e, \u003cem\u003e2016\u003c/em\u003e, 1\u0026ndash;14. https://doi.org/10.1155/2016/5685970\u003c/li\u003e\n\u003cli\u003eHa, J., \u0026amp; Lee, J.-S. (2016). A New Under-Sampling Method Using Genetic Algorithm for Imbalanced Data Classification. \u003cem\u003eProceedings of the 10th International Conference on Ubiquitous Information Management and Communication\u003c/em\u003e, 1\u0026ndash;6. https://doi.org/10.1145/2857546.2857643\u003c/li\u003e\n\u003cli\u003eHaibo He, Yang Bai, Garcia, E. A., \u0026amp; Shutao Li. (2008). ADASYN: Adaptive synthetic sampling approach for imbalanced learning. \u003cem\u003e2008 IEEE International Joint Conference on Neural Networks (IEEE World Congress on Computational Intelligence)\u003c/em\u003e, 1322\u0026ndash;1328. https://doi.org/10.1109/IJCNN.2008.4633969\u003c/li\u003e\n\u003cli\u003eHancock, J., Khoshgoftaar, T. M., \u0026amp; Johnson, J. M. (2022). The Effects of Random Undersampling for Big Data Medicare Fraud Detection. \u003cem\u003e2022 IEEE International Conference on Service-Oriented System Engineering (SOSE)\u003c/em\u003e, 141\u0026ndash;146. https://doi.org/10.1109/SOSE55356.2022.00023\u003c/li\u003e\n\u003cli\u003eHancock, J. T., \u0026amp; Khoshgoftaar, T. M. (2020). CatBoost for big data: An interdisciplinary review. \u003cem\u003eJournal of Big Data\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e(1), 94. https://doi.org/10.1186/s40537-020-00369-8\u003c/li\u003e\n\u003cli\u003eHasanin, T., Khoshgoftaar, T. M., Leevy, J. L., \u0026amp; Bauder, R. A. (2019). Severely imbalanced Big Data challenges: Investigating data sampling approaches. \u003cem\u003eJournal of Big Data\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e(1), 107. https://doi.org/10.1186/s40537-019-0274-4\u003c/li\u003e\n\u003cli\u003eHernandez, J., Carrasco-Ochoa, J. A., \u0026amp; Mart\u0026iacute;nez-Trinidad, J. F. (2013). An Empirical Study of Oversampling and Undersampling for Instance Selection Methods on Imbalance Datasets. In J. Ruiz-Shulcloper \u0026amp; G. Sanniti di Baja (Eds.), \u003cem\u003eProgress in Pattern Recognition, Image Analysis, Computer Vision, and Applications\u003c/em\u003e (Vol. 8258, pp. 262\u0026ndash;269). Springer Berlin Heidelberg. https://doi.org/10.1007/978-3-642-41822-8_33\u003c/li\u003e\n\u003cli\u003eHooda, N., Bawa, S., \u0026amp; Rana, P. S. (2018). Fraudulent Firm Classification: A Case Study of an External Audit. \u003cem\u003eApplied Artificial Intelligence\u003c/em\u003e, \u003cem\u003e32\u003c/em\u003e(1), 48\u0026ndash;64. https://doi.org/10.1080/08839514.2018.1451032\u003c/li\u003e\n\u003cli\u003eHuan, W., Lin, H., Li, H., Zhou, Y., \u0026amp; Wang, Y. (2020). Anomaly Detection Method Based on Clustering Undersampling and Ensemble Learning. \u003cem\u003e2020 IEEE 5th Information Technology and Mechatronics Engineering Conference (ITOEC)\u003c/em\u003e, 980\u0026ndash;984. https://doi.org/10.1109/ITOEC49072.2020.9141897\u003c/li\u003e\n\u003cli\u003eHuang, G., Wu, L., Ma, X., Zhang, W., Fan, J., Yu, X., Zeng, W., \u0026amp; Zhou, H. (2019). Evaluation of CatBoost method for prediction of reference evapotranspiration in humid regions. \u003cem\u003eJournal of Hydrology\u003c/em\u003e, \u003cem\u003e574\u003c/em\u003e, 1029\u0026ndash;1041. https://doi.org/10.1016/j.jhydrol.2019.04.085\u003c/li\u003e\n\u003cli\u003eIjaz, M. F., Attique, M., \u0026amp; Son, Y. (2020). Data-Driven Cervical Cancer Prediction Model with Outlier Detection and Over-Sampling Methods. \u003cem\u003eSensors\u003c/em\u003e, \u003cem\u003e20\u003c/em\u003e(10), 2809. https://doi.org/10.3390/s20102809\u003c/li\u003e\n\u003cli\u003eJeni, L. A., Cohn, J. F., \u0026amp; De La Torre, F. (2013). Facing Imbalanced Data\u0026mdash;Recommendations for the Use of Performance Metrics. \u003cem\u003e2013 Humaine Association Conference on Affective Computing and Intelligent Interaction\u003c/em\u003e, 245\u0026ndash;251. https://doi.org/10.1109/ACII.2013.47\u003c/li\u003e\n\u003cli\u003eJo, T., \u0026amp; Japkowicz, N. (2004). Class imbalances versus small disjuncts. \u003cem\u003eACM SIGKDD Explorations Newsletter\u003c/em\u003e, \u003cem\u003e6\u003c/em\u003e(1), 40\u0026ndash;49. https://doi.org/10.1145/1007730.1007737\u003c/li\u003e\n\u003cli\u003eKamei, Y., Monden, A., Matsumoto, S., Kakimoto, T., \u0026amp; Matsumoto, K. (2007). The Effects of Over and Under Sampling on Fault-prone Module Detection. \u003cem\u003eFirst International Symposium on Empirical Software Engineering and Measurement (ESEM 2007)\u003c/em\u003e, 196\u0026ndash;204. https://doi.org/10.1109/ESEM.2007.28\u003c/li\u003e\n\u003cli\u003eKeller, A., \u0026amp; Pandey, A. (2021). SMOTE and ENN based XGBoost prediction model for Parkinson\u0026rsquo;s disease detection. \u003cem\u003e2021 2nd International Conference on Smart Electronics and Communication (ICOSEC)\u003c/em\u003e, 839\u0026ndash;846. https://doi.org/10.1109/ICOSEC51865.2021.9591716\u003c/li\u003e\n\u003cli\u003eKhan, T. M., Xu, S., Khan, Z. G., \u0026amp; Uzair chishti, M. (2021). Implementing Multilabeling, ADASYN, and ReliefF Techniques for Classification of Breast Cancer Diagnostic through Machine Learning: Efficient Computer-Aided Diagnostic System. \u003cem\u003eJournal of Healthcare Engineering\u003c/em\u003e, \u003cem\u003e2021\u003c/em\u003e, 1\u0026ndash;15. https://doi.org/10.1155/2021/5577636\u003c/li\u003e\n\u003cli\u003eKhushi, M., Shaukat, K., Alam, T. M., Hameed, I. A., Uddin, S., Luo, S., Yang, X., \u0026amp; Reyes, M. C. (2021). A Comparative Performance Analysis of Data Resampling Methods on Imbalance Medical Data. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e9\u003c/em\u003e, 109960\u0026ndash;109975. https://doi.org/10.1109/ACCESS.2021.3102399\u003c/li\u003e\n\u003cli\u003eKrayem, A., Yeretzian, A., Faour, G., \u0026amp; Najem, S. (2021). Machine learning for buildings\u0026rsquo; characterization and power-law recovery of urban metrics. \u003cem\u003ePLOS ONE\u003c/em\u003e, \u003cem\u003e16\u003c/em\u003e(1), e0246096. https://doi.org/10.1371/journal.pone.0246096\u003c/li\u003e\n\u003cli\u003eKurani, A., Doshi, P., Vakharia, A., \u0026amp; Shah, M. (2023). A Comprehensive Comparative Study of Artificial Neural Network (ANN) and Support Vector Machines (SVM) on Stock Forecasting. \u003cem\u003eAnnals of Data Science\u003c/em\u003e, \u003cem\u003e10\u003c/em\u003e(1), 183\u0026ndash;208. https://doi.org/10.1007/s40745-021-00344-x\u003c/li\u003e\n\u003cli\u003eKurniawati, Y. E., Permanasari, A. E., \u0026amp; Fauziati, S. (2018). Adaptive Synthetic-Nominal (ADASYN-N) and Adaptive Synthetic-KNN (ADASYN-KNN) for Multiclass Imbalance Learning on Laboratory Test Data. \u003cem\u003e2018 4th International Conference on Science and Technology (ICST)\u003c/em\u003e, 1\u0026ndash;6. https://doi.org/10.1109/ICSTC.2018.8528679\u003c/li\u003e\n\u003cli\u003eLamari, M., Azizi, N., Hammami, N. E., Boukhamla, A., Cheriguene, S., Dendani, N., \u0026amp; Benzebouchi, N. E. (2021). SMOTE\u0026ndash;ENN-Based Data Sampling and Improved Dynamic Ensemble Selection for Imbalanced Medical Data Classification. In F. Saeed, T. Al-Hadhrami, F. Mohammed, \u0026amp; E. Mohammed (Eds.), \u003cem\u003eAdvances on Smart and Soft Computing\u003c/em\u003e (Vol. 1188, pp. 37\u0026ndash;49). Springer Singapore. https://doi.org/10.1007/978-981-15-6048-4_4\u003c/li\u003e\n\u003cli\u003eLin, C., Tsai, C.-F., \u0026amp; Lin, W.-C. (2023). Towards hybrid over- and under-sampling combination methods for class imbalanced datasets: An experimental study. \u003cem\u003eArtificial Intelligence Review\u003c/em\u003e, \u003cem\u003e56\u003c/em\u003e(2), 845\u0026ndash;863. https://doi.org/10.1007/s10462-022-10186-5\u003c/li\u003e\n\u003cli\u003eLiu, R., Huang, J., \u0026amp; Zhang, Z. (2023). Tracking disclosure change trajectories for financial fraud detection. \u003cem\u003eProduction and Operations Management\u003c/em\u003e, \u003cem\u003e32\u003c/em\u003e(2), 584\u0026ndash;602. https://doi.org/10.1111/poms.13888\u003c/li\u003e\n\u003cli\u003eLokanan, M. E. (2023). Predicting mobile money transaction fraud using machine learning algorithms. \u003cem\u003eApplied AI Letters\u003c/em\u003e, \u003cem\u003e4\u003c/em\u003e(2), e85. https://doi.org/10.1002/ail2.85\u003c/li\u003e\n\u003cli\u003eLokanan, M. E., \u0026amp; Sharma, K. (2022). Fraud prediction using machine learning: The case of investment advisors in Canada. \u003cem\u003eMachine Learning with Applications\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e, 100269. https://doi.org/10.1016/j.mlwa.2022.100269\u003c/li\u003e\n\u003cli\u003eLokanan, M., \u0026amp; Sharma, S. (2024). The use of machine learning algorithms to predict financial statement fraud. \u003cem\u003eThe British Accounting Review\u003c/em\u003e, 101441. https://doi.org/10.1016/j.bar.2024.101441\u003c/li\u003e\n\u003cli\u003eLu, C., Lin, S., Liu, X., \u0026amp; Shi, H. (2020). Telecom Fraud Identification Based on ADASYN and Random Forest. \u003cem\u003e2020 5th International Conference on Computer and Communication Systems (ICCCS)\u003c/em\u003e, 447\u0026ndash;452. https://doi.org/10.1109/ICCCS49078.2020.9118521\u003c/li\u003e\n\u003cli\u003eMaldonado, S., L\u0026oacute;pez, J., \u0026amp; Vairetti, C. (2019). An alternative SMOTE oversampling strategy for high-dimensional datasets. \u003cem\u003eApplied Soft Computing\u003c/em\u003e, \u003cem\u003e76\u003c/em\u003e, 380\u0026ndash;389. https://doi.org/10.1016/j.asoc.2018.12.024\u003c/li\u003e\n\u003cli\u003eMienye, I. D., \u0026amp; Sun, Y. (2023). A Deep Learning Ensemble With Data Resampling for Credit Card Fraud Detection. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e, 30628\u0026ndash;30638. https://doi.org/10.1109/ACCESS.2023.3262020\u003c/li\u003e\n\u003cli\u003eMore, A. (2016). \u003cem\u003eSurvey of resampling techniques for improving classification performance in unbalanced datasets\u003c/em\u003e. https://doi.org/10.48550/ARXIV.1608.06048\u003c/li\u003e\n\u003cli\u003eMoreo, A., Esuli, A., \u0026amp; Sebastiani, F. (2016). Distributional Random Oversampling for Imbalanced Text Classification. \u003cem\u003eProceedings of the 39th International ACM SIGIR Conference on Research and Development in Information Retrieval\u003c/em\u003e, 805\u0026ndash;808. https://doi.org/10.1145/2911451.2914722\u003c/li\u003e\n\u003cli\u003eMqadi, N. M., Naicker, N., \u0026amp; Adeliyi, T. (2021). Solving Misclassification of the Credit Card Imbalance Problem Using Near Miss. \u003cem\u003eMathematical Problems in Engineering\u003c/em\u003e, \u003cem\u003e2021\u003c/em\u003e, 1\u0026ndash;16. https://doi.org/10.1155/2021/7194728\u003c/li\u003e\n\u003cli\u003eMunshi, R. M. (2024). Correction: Novel ensemble learning approach with SVM-imputed ADASYN features for enhanced cervical cancer prediction. \u003cem\u003ePLOS ONE\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e(2), e0298980. https://doi.org/10.1371/journal.pone.0298980\u003c/li\u003e\n\u003cli\u003eMuranda, C., Ali, A., \u0026amp; Shongwe, T. (2020). Detecting Fraudulent Motor Insurance Claims Using Support Vector Machines with Adaptive Synthetic Sampling Method. \u003cem\u003e2020 61st International Scientific Conference on Information Technology and Management Science of Riga Technical University (ITMS)\u003c/em\u003e, 1\u0026ndash;5. https://doi.org/10.1109/ITMS51158.2020.9259322\u003c/li\u003e\n\u003cli\u003eNaseriparsa, M., \u0026amp; Kashani, M. M. R. (2014). \u003cem\u003eCombination of PCA with SMOTE Resampling to Boost the Prediction Rate in Lung Cancer Dataset\u003c/em\u003e. https://doi.org/10.48550/ARXIV.1403.1949\u003c/li\u003e\n\u003cli\u003eNayan, N. M., Islam, A., Islam, M. U., Ahmed, E., Hossain, M. M., \u0026amp; Alam, M. Z. (2023). SMOTE Oversampling and Near Miss Undersampling Based Diabetes Diagnosis from Imbalanced Dataset with XAI Visualization. \u003cem\u003e2023 IEEE Symposium on Computers and Communications (ISCC)\u003c/em\u003e, 1\u0026ndash;6. https://doi.org/10.1109/ISCC58397.2023.10218281\u003c/li\u003e\n\u003cli\u003eNguyen, H. M., Cooper, E. W., \u0026amp; Kamei, K. (2011). Borderline over-sampling for imbalanced data classification. \u003cem\u003eInternational Journal of Knowledge Engineering and Soft Data Paradigms\u003c/em\u003e, \u003cem\u003e3\u003c/em\u003e(1), 4. https://doi.org/10.1504/IJKESDP.2011.039875\u003c/li\u003e\n\u003cli\u003eNing, Q., Zhao, X., \u0026amp; Ma, Z. (2022). A Novel Method for Identification of Glutarylation Sites Combining Borderline-SMOTE With Tomek Links Technique in Imbalanced Data. \u003cem\u003eIEEE/ACM Transactions on Computational Biology and Bioinformatics\u003c/em\u003e, \u003cem\u003e19\u003c/em\u003e(5), 2632\u0026ndash;2641. https://doi.org/10.1109/TCBB.2021.3095482\u003c/li\u003e\n\u003cli\u003eOladunni, T., Tossou, S., Haile, Y., \u0026amp; Kidane, A. (2021). \u003cem\u003eCOVID-19 County Level Severity Classification with Imbalanced Dataset: A NearMiss Under-sampling Approach\u003c/em\u003e [Preprint]. Epidemiology. https://doi.org/10.1101/2021.05.21.21257603\u003c/li\u003e\n\u003cli\u003ePan, T., Zhao, J., Wu, W., \u0026amp; Yang, J. (2020). Learning imbalanced datasets based on SMOTE and Gaussian distribution. \u003cem\u003eInformation Sciences\u003c/em\u003e, \u003cem\u003e512\u003c/em\u003e, 1214\u0026ndash;1233. https://doi.org/10.1016/j.ins.2019.10.048\u003c/li\u003e\n\u003cli\u003ePereira, R. M., Costa, Y. M. G., \u0026amp; Silla Jr., C. N. (2020). MLTL: A multi-label approach for the Tomek Link undersampling algorithm. \u003cem\u003eNeurocomputing\u003c/em\u003e, \u003cem\u003e383\u003c/em\u003e, 95\u0026ndash;105. https://doi.org/10.1016/j.neucom.2019.11.076\u003c/li\u003e\n\u003cli\u003ePias, T. S., Su, Y., Tang, X., Wang, H., Faghani, S., \u0026amp; Yao, D. (Daphne). (2023). \u003cem\u003eEnhancing Fairness and Accuracy in Diagnosing Type 2 Diabetes in Young Population\u003c/em\u003e [Preprint]. Health Informatics. https://doi.org/10.1101/2023.05.02.23289405\u003c/li\u003e\n\u003cli\u003eRajer-Kanduč, K., Zupan, J., \u0026amp; Majcen, N. (2003). Separation of data on the training and test set for modelling: A case study for modelling of five colour properties of a white pigment. \u003cem\u003eChemometrics and Intelligent Laboratory Systems\u003c/em\u003e, \u003cem\u003e65\u003c/em\u003e(2), 221\u0026ndash;229. https://doi.org/10.1016/S0169-7439(02)00110-7\u003c/li\u003e\n\u003cli\u003eRao, C., Xu, Y., Xiao, X., Hu, F., \u0026amp; Goh, M. (2024). Imbalanced customer churn classification using a new multi-strategy collaborative processing method. \u003cem\u003eExpert Systems with Applications\u003c/em\u003e, \u003cem\u003e247\u003c/em\u003e, 123251. https://doi.org/10.1016/j.eswa.2024.123251\u003c/li\u003e\n\u003cli\u003eRubaidi, Z. S., Ammar, B. B., \u0026amp; Aouicha, M. B. (2022). Fraud Detection Using Large-scale Imbalance Dataset. \u003cem\u003eInternational Journal on Artificial Intelligence Tools\u003c/em\u003e, \u003cem\u003e31\u003c/em\u003e(08), 2250037. https://doi.org/10.1142/S0218213022500373\u003c/li\u003e\n\u003cli\u003eSantos, M. S., Soares, J. P., Abreu, P. H., Araujo, H., \u0026amp; Santos, J. (2018). Cross-Validation for Imbalanced Datasets: Avoiding Overoptimistic and Overfitting Approaches [Research Frontier]. \u003cem\u003eIEEE Computational Intelligence Magazine\u003c/em\u003e, \u003cem\u003e13\u003c/em\u003e(4), 59\u0026ndash;76. https://doi.org/10.1109/MCI.2018.2866730\u003c/li\u003e\n\u003cli\u003eSaripuddin, M., Suliman, A., Syarmila Sameon, S., \u0026amp; Jorgensen, B. N. (2021). Random Undersampling on Imbalance Time Series Data for Anomaly Detection. \u003cem\u003e2021 The 4th International Conference on Machine Learning and Machine Intelligence\u003c/em\u003e, 151\u0026ndash;156. https://doi.org/10.1145/3490725.3490748\u003c/li\u003e\n\u003cli\u003eSilva, B., Silveira, R., Silva Neto, M., Cortez, P., \u0026amp; Gomes, D. (2021). A comparative analysis of undersampling techniques for network intrusion detection systems design. \u003cem\u003eJournal of Communication and Information Systems\u003c/em\u003e, \u003cem\u003e36\u003c/em\u003e(1), 31\u0026ndash;43. https://doi.org/10.14209/jcis.2021.3\u003c/li\u003e\n\u003cli\u003eSingh, A., Ranjan, R. K., \u0026amp; Tiwari, A. (2022). Credit Card Fraud Detection under Extreme Imbalanced Data: A Comparative Study of Data-level Algorithms. \u003cem\u003eJournal of Experimental \u0026amp; Theoretical Artificial Intelligence\u003c/em\u003e, \u003cem\u003e34\u003c/em\u003e(4), 571\u0026ndash;598. https://doi.org/10.1080/0952813X.2021.1907795\u003c/li\u003e\n\u003cli\u003eSisodia, D. S., Reddy, N. K., \u0026amp; Bhandari, S. (2017). Performance evaluation of class balancing techniques for credit card fraud detection. \u003cem\u003e2017 IEEE International Conference on Power, Control, Signals and Instrumentation Engineering (ICPCSI)\u003c/em\u003e, 2747\u0026ndash;2752. https://doi.org/10.1109/ICPCSI.2017.8392219\u003c/li\u003e\n\u003cli\u003eSrinilta, C., \u0026amp; Kanharattanachai, S. (2021). Application of Natural Neighbor-based Algorithm on Oversampling SMOTE Algorithms. \u003cem\u003e2021 7th International Conference on Engineering, Applied Sciences and Technology (ICEAST)\u003c/em\u003e, 217\u0026ndash;220. https://doi.org/10.1109/ICEAST52143.2021.9426310\u003c/li\u003e\n\u003cli\u003eSrinivasan, S., Vallikannu, A. L., Manoharan, L., Deepthi, K., \u0026amp; Aravind Yadav, B. (2024). Identification of the Best Combination of Oversampling Technique and Machine Learning Algorithm for Credit Card Fraud Detection. In I. J. Jacob, S. Piramuthu, \u0026amp; P. Falkowski-Gilski (Eds.), \u003cem\u003eData Intelligence and Cognitive Informatics\u003c/em\u003e (pp. 557\u0026ndash;571). Springer Nature Singapore. https://doi.org/10.1007/978-981-99-7962-2_41\u003c/li\u003e\n\u003cli\u003eSubudhi, S., \u0026amp; Panigrahi, S. (2018). Effect of Class Imbalanceness in Detecting Automobile Insurance Fraud. \u003cem\u003e2018 2nd International Conference on Data Science and Business Analytics (ICDSBA)\u003c/em\u003e, 528\u0026ndash;531. https://doi.org/10.1109/ICDSBA.2018.00104\u003c/li\u003e\n\u003cli\u003eSujitha, R., \u0026amp; Paramasivan, B. (2023). Optimal progressive classification study using SMOTE-SVM for stages of lung disease. \u003cem\u003eAutomatika\u003c/em\u003e, \u003cem\u003e64\u003c/em\u003e(4), 807\u0026ndash;814. https://doi.org/10.1080/00051144.2023.2218167\u003c/li\u003e\n\u003cli\u003eSun, B., \u0026amp; Chen, H. (2021). A Survey of k Nearest Neighbor Algorithms for Solving the Class Imbalanced Problem. \u003cem\u003eWireless Communications and Mobile Computing\u003c/em\u003e, \u003cem\u003e2021\u003c/em\u003e, 1\u0026ndash;12. https://doi.org/10.1155/2021/5520990\u003c/li\u003e\n\u003cli\u003eSun, T., \u0026amp; Vasarhelyi, M. A. (2018). Predicting credit card delinquencies: An application of deep neural networks. \u003cem\u003eIntelligent Systems in Accounting, Finance and Management\u003c/em\u003e, \u003cem\u003e25\u003c/em\u003e(4), 174\u0026ndash;189. https://doi.org/10.1002/isaf.1437\u003c/li\u003e\n\u003cli\u003eTanimu, J. J., Hamada, M., Hassan, M., Kakudi, H., \u0026amp; Abiodun, J. O. (2022). A Machine Learning Method for Classification of Cervical Cancer. \u003cem\u003eElectronics\u003c/em\u003e, \u003cem\u003e11\u003c/em\u003e(3), 463. https://doi.org/10.3390/electronics11030463\u003c/li\u003e\n\u003cli\u003eTariq, M. A., Sargano, A. B., Iftikhar, M. A., \u0026amp; Habib, Z. (2023). Comparing Different Oversampling Methods in Predicting Multi-Class Educational Datasets Using Machine Learning Techniques. \u003cem\u003eCybernetics and Information Technologies\u003c/em\u003e, \u003cem\u003e23\u003c/em\u003e(4), 199\u0026ndash;212. https://doi.org/10.2478/cait-2023-0044\u003c/li\u003e\n\u003cli\u003eTiwari, A. K., Nath, A., Subbiah, K., \u0026amp; Shukla, K. K. (2015). Effect of varying degree of resampling on prediction accuracy for observed peptide count in protein mass spectrometry data. \u003cem\u003e2015 11th International Conference on Natural Computation (ICNC)\u003c/em\u003e, 691\u0026ndash;695. https://doi.org/10.1109/ICNC.2015.7378074\u003c/li\u003e\n\u003cli\u003eTiwari, S., Wee, H. M., \u0026amp; Daryanto, Y. (2018). Big data analytics in supply chain management between 2010 and 2016: Insights to industries. \u003cem\u003eComputers \u0026amp; Industrial Engineering\u003c/em\u003e, \u003cem\u003e115\u003c/em\u003e, 319\u0026ndash;330. https://doi.org/10.1016/j.cie.2017.11.017\u003c/li\u003e\n\u003cli\u003eVuttipittayamongkol, P., Elyan, E., \u0026amp; Petrovski, A. (2021). On the class overlap problem in imbalanced data classification. \u003cem\u003eKnowledge-Based Systems\u003c/em\u003e, \u003cem\u003e212\u003c/em\u003e, 106631. https://doi.org/10.1016/j.knosys.2020.106631\u003c/li\u003e\n\u003cli\u003eWang, K., Tian, J., Zheng, C., Yang, H., Ren, J., Li, C., Han, Q., \u0026amp; Zhang, Y. (2021). Improving Risk Identification of Adverse Outcomes in Chronic Heart Failure Using SMOTE+ENN and Machine Learning. \u003cem\u003eRisk Management and Healthcare Policy\u003c/em\u003e, \u003cem\u003eVolume 14\u003c/em\u003e, 2453\u0026ndash;2463. https://doi.org/10.2147/RMHP.S310295\u003c/li\u003e\n\u003cli\u003eWang, Z., Wu, C., Zheng, K., Niu, X., \u0026amp; Wang, X. (2019). SMOTETomek-Based Resampling for Personality Recognition. \u003cem\u003eIEEE Access\u003c/em\u003e, \u003cem\u003e7\u003c/em\u003e, 129678\u0026ndash;129689. https://doi.org/10.1109/ACCESS.2019.2940061\u003c/li\u003e\n\u003cli\u003eXie, X., Liu, H., Zeng, S., Lin, L., \u0026amp; Li, W. (2021). A novel progressively undersampling method based on the density peaks sequence for imbalanced data. \u003cem\u003eKnowledge-Based Systems\u003c/em\u003e, \u003cem\u003e213\u003c/em\u003e, 106689. https://doi.org/10.1016/j.knosys.2020.106689\u003c/li\u003e\n\u003cli\u003eXu, Z., Shen, D., Nie, T., \u0026amp; Kou, Y. (2020). A hybrid sampling algorithm combining M-SMOTE and ENN based on Random forest for medical imbalanced data. \u003cem\u003eJournal of Biomedical Informatics\u003c/em\u003e, \u003cem\u003e107\u003c/em\u003e, 103465. https://doi.org/10.1016/j.jbi.2020.103465\u003c/li\u003e\n\u003cli\u003eXu, Z., Shen, D., Nie, T., Kou, Y., Yin, N., \u0026amp; Han, X. (2021). A cluster-based oversampling algorithm combining SMOTE and k-means for imbalanced medical data. \u003cem\u003eInformation Sciences\u003c/em\u003e, \u003cem\u003e572\u003c/em\u003e, 574\u0026ndash;589. https://doi.org/10.1016/j.ins.2021.02.056\u003c/li\u003e\n\u003cli\u003eXuan, S., Liu, G., Li, Z., Zheng, L., Wang, S., \u0026amp; Jiang, C. (2018). Random forest for credit card fraud detection. \u003cem\u003e2018 IEEE 15th International Conference on Networking, Sensing and Control (ICNSC)\u003c/em\u003e, 1\u0026ndash;6. https://doi.org/10.1109/ICNSC.2018.8361343\u003c/li\u003e\n\u003cli\u003eYang, F., Wang, K., Sun, L., Zhai, M., Song, J., \u0026amp; Wang, H. (2022). A hybrid sampling algorithm combining synthetic minority over-sampling technique and edited nearest neighbor for missed abortion diagnosis. \u003cem\u003eBMC Medical Informatics and Decision Making\u003c/em\u003e, \u003cem\u003e22\u003c/em\u003e(1), 344. https://doi.org/10.1186/s12911-022-02075-2\u003c/li\u003e\n\u003cli\u003eYang, Y., Yang, X., Tang, W., \u0026amp; Li, L. (2023). A Undersampling-DoppelGANger based Data Generation Method for Unbalanced BGP Data. \u003cem\u003e2023 IEEE 9th International Conference on Cloud Computing and Intelligent Systems (CCIS)\u003c/em\u003e, 100\u0026ndash;105. https://doi.org/10.1109/CCIS59572.2023.10263221\u003c/li\u003e\n\u003cli\u003eYap, B. W., Rani, K. A., Rahman, H. A. A., Fong, S., Khairudin, Z., \u0026amp; Abdullah, N. N. (2014). An Application of Oversampling, Undersampling, Bagging and Boosting in Handling Imbalanced Datasets. In T. Herawan, M. M. Deris, \u0026amp; J. Abawajy (Eds.), \u003cem\u003eProceedings of the First International Conference on Advanced Data and Information Engineering (DaEng-2013)\u003c/em\u003e (Vol. 285, pp. 13\u0026ndash;22). Springer Singapore. https://doi.org/10.1007/978-981-4585-18-7_2\u003c/li\u003e\n\u003cli\u003eZeng, M., Zou, B., Wei, F., Liu, X., \u0026amp; Wang, L. (2016). Effective prediction of three common diseases by combining SMOTE with Tomek links technique for imbalanced medical data. \u003cem\u003e2016 IEEE International Conference of Online Analysis and Computing Science (ICOACS)\u003c/em\u003e, 225\u0026ndash;228. https://doi.org/10.1109/ICOACS.2016.7563084\u003c/li\u003e\n\u003cli\u003eZhang, H., \u0026amp; Li, M. (2014). RWO-Sampling: A random walk over-sampling approach to imbalanced data classification. \u003cem\u003eInformation Fusion\u003c/em\u003e, \u003cem\u003e20\u003c/em\u003e, 99\u0026ndash;116. https://doi.org/10.1016/j.inffus.2013.12.003\u003c/li\u003e\n\u003cli\u003eZhao, H., Chen, X., Nguyen, T., Huang, J. Z., Williams, G., \u0026amp; Chen, H. (2016). Stratified Over-Sampling Bagging Method for Random Forests on Imbalanced Data. In M. Chau, G. A. Wang, \u0026amp; H. Chen (Eds.), \u003cem\u003eIntelligence and Security Informatics\u003c/em\u003e (Vol. 9650, pp. 63\u0026ndash;72). Springer International Publishing. https://doi.org/10.1007/978-3-319-31863-9_5\u003c/li\u003e\n\u003cli\u003eZhou, L. (2013). Performance of corporate bankruptcy prediction models on imbalanced dataset: The effect of sampling methods. \u003cem\u003eKnowledge-Based Systems\u003c/em\u003e, \u003cem\u003e41\u003c/em\u003e, 16\u0026ndash;25. https://doi.org/10.1016/j.knosys.2012.12.007\u003c/li\u003e\n\u003cli\u003eZhu, T., Lin, Y., \u0026amp; Liu, Y. (2017). Synthetic minority oversampling technique for multiclass imbalance problems. \u003cem\u003ePattern Recognition\u003c/em\u003e, \u003cem\u003e72\u003c/em\u003e, 327\u0026ndash;340. https://doi.org/10.1016/j.patcog.2017.07.024\u003c/li\u003e\n\u003cli\u003eZuech, R., Hancock, J., \u0026amp; Khoshgoftaar, T. M. (2021). Detecting web attacks using random undersampling and ensemble learners. \u003cem\u003eJournal of Big Data\u003c/em\u003e, \u003cem\u003e8\u003c/em\u003e(1), 75. https://doi.org/10.1186/s40537-021-00460-8\u003c/li\u003e\n\u003c/ol\u003e"},{"header":"Footnotes","content":"\u003col\u003e\u003cli\u003e\u003cspan\u003e Yeh,I-Cheng. (2016). Default of credit card clients. UCI Machine Learning Repository. \u003cspan class=\"ExternalRef\"\u003e\u003cspan class=\"RefSource\"\u003ehttps://doi.org/10.24432/C55S3H\u003c/span\u003e\u003cspan address=\"10.24432/C55S3H\" targettype=\"DOI\" class=\"RefTarget\"\u003e\u003c/span\u003e\u003c/span\u003e.\u003c/span\u003e\u003c/li\u003e\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"Royal Roads University","isAcceptedByJournal":false,"isAuthorSuppliedPdf":false,"isDeskRejected":"","isHiddenFromSearch":false,"isInQc":false,"isInWorkflow":false,"isPdf":false,"isPdfUpToDate":true,"isWithdrawnOrRetracted":false,"journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true},"keywords":"Undersampling, Oversampling, Machine learning, Credit Default, Loan Default, DART, Imbalance dataset","lastPublishedDoi":"10.21203/rs.3.rs-6895500/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-6895500/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eClass imbalance remains a persistent challenge in predictive modeling, often leading to biased machine learning outcomes that disproportionately favor the majority class. This study investigates the effectiveness of advanced resampling techniques\u0026mdash;both undersampling and oversampling\u0026mdash;across two large and highly imbalanced datasets involving credit and loan default prediction. In addition to evaluating established oversampling techniques, the study introduces and validates a novel resampling approach, DART (Deep Adaptive Resampling Technique). Each technique is assessed using a consistent suite of classifiers, including logistic regression, gradient descent, Na\u0026iuml;ve Bayes, random forest, CatBoost, and artificial neural networks. The results reveal that K-MeansSMOTE and NearMiss outperform other resampling strategies in oversampling and undersampling, respectively, by achieving balanced trade-offs in precision, recall, F1-score, AUC, and Matthews Correlation Coefficient. Notably, DART demonstrates exceptional performance across both datasets, achieving nearly perfect classification scores across all metrics, suggesting strong generalizability and robustness. The study further analyzes the strengths and limitations of each resampling technique and emphasizes the importance of metric selection when evaluating imbalanced datasets. By integrating empirical evaluation with theoretical insights, this research contributes to the growing body of literature on imbalanced learning and offers practical guidance for selecting appropriate resampling strategies. These findings have broader implications for domains such as finance, healthcare, and fraud detection, where class imbalance is common. Overall, the study affirms the value of hybrid and adaptive resampling methods in building more accurate and generalizable predictive models.\u003c/p\u003e","manuscriptTitle":"Introducing DART: A Novel Deep Adaptive Upsampling Technique for Handling Class Imbalance","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-06-18 08:45:49","doi":"10.21203/rs.3.rs-6895500/v1","editorialEvents":[{"type":"communityComments","content":0}],"status":"published","journal":{"display":true,"email":"[email protected]","identity":"researchsquare","isNatureJournal":false,"hasQc":true,"allowDirectSubmit":true,"externalIdentity":"","sideBox":"","snPcode":"","submissionUrl":"/submission","title":"Research Square","twitterHandle":"researchsquare","acdcEnabled":true,"dfaEnabled":false,"editorialSystem":"","reportingPortfolio":"","inReviewEnabled":false,"inReviewRevisionsEnabled":true}}],"origin":"","ownerIdentity":"144c5219-61aa-47e6-ae8d-6a6c2503d85d","owner":[],"postedDate":"June 18th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[{"id":50056132,"name":"Artificial Intelligence and Machine Learning"}],"tags":[],"updatedAt":"2025-06-18T08:45:50+00:00","versionOfRecord":[],"versionCreatedAt":"2025-06-18 08:45:49","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-6895500","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-6895500","identity":"rs-6895500","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}