Employee Attrition Prediction System using Machine Learning and Artificial Intelligence

Employee Attrition Prediction System using Machine Learning and Artificial Intelligence | 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 Employee Attrition Prediction System using Machine Learning and Artificial Intelligence Anurag Bodkhe, Sahil Jirapure, Ujjwal Garud, Shrinivas Bhore This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-9141427/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 Employee attrition remains one of the most consequential workforce challenges facing contemporary organizations, with replacement costs estimated between 50% and 200% of an affected employee’s annual compensation. This paper presents the design, implementation, and empirical evaluation of an Employee Attrition Prediction System (EAPS) built on supervised machine learning techniques applied to the IBM HR Analytics dataset comprising 1,470 employee records and 35 workforce attributes. Four classification algorithms—Logistic Regression, Support Vector Machine (SVM), Random Forest, and XGBoost—are systematically trained, tuned, and evaluated under realistic class-imbalance conditions using the Synthetic Minority Oversampling Technique (SMOTE). Three domain-informed engineered features are introduced to augment the base feature set: Compensation Ratio, Tenure per Job, and Years Without Change. Experimental results demonstrate that XGBoost achieves superior performance across all five evaluation metrics, attaining 97.2% accuracy, 96.8% precision, 95.4% recall, a macro F1 score of 96.1%, and an AUC-ROC of 0.991 following stratified 10-fold cross-validation and hyperparameter optimization. A modular six-component system architecture is proposed, culminating in an HR decision-support dashboard that leverages SHAP (SHapley Additive exPlanations) values to deliver individualized, interpretable attrition risk assessments to non-technical HR practitioners. The proposed system addresses the critical gap between available HR data and proactive workforce retention strategy, providing organizations with a scalable, evidence-based tool for reducing voluntary turnover and its associated organizational costs. Machine Learning Human Resource Analytics Employee Attrition Prediction XGBoost Random Forest SMOTE Explainable AI Binary Classification Workforce Management IBM HR Dataset SHAP Feature Engineering I. INTRODUCTION Employee attrition—defined as the voluntary resignation of an employee from an organization—constitutes a recurring and financially significant problem that affects enterprises across every industry sector. When a trained and experienced worker departs, the organization absorbs multiple categories of cost that extend well beyond the immediate expense of posting a job vacancy. Direct costs include recruiter fees, advertising expenditures, interview panel time, background verification, onboarding programs, and the reduced productivity of the incoming employee during the initial learning period. Indirect costs, which are harder to quantify but no less consequential, encompass the erosion of project continuity, the loss of client relationships cultivated by the departing employee, and the motivational impact on the remaining workforce who may interpret a colleague’s departure as a signal of deteriorating organizational conditions. Estimates from the Society for Human Resource Management (SHRM) and independent academic research consistently indicate that the total cost of replacing a mid-level employee ranges from 50% to 200% of that employee’s annual salary. For senior technical and managerial roles, this figure can exceed 300%. When aggregated across an organization of thousands of employees experiencing even modest annual turnover rates of 15–20%, the cumulative financial burden becomes one of the largest controllable operational expenditures in HR management. Despite the scale of this challenge, the dominant paradigm in HR practice remains reactive. Exit interviews and post-departure surveys yield retrospective insights that arrive too late to prevent the resignation they seek to understand. Annual engagement surveys provide periodic snapshots that fail to capture the granular, employee-level trajectory of disengagement that typically precedes voluntary departure by months. What modern organizations require is a forward-looking, continuously updated risk assessment system that can identify individual employees who are approaching a resignation decision while there is still sufficient lead time for targeted retention interventions. Advances in machine learning and the proliferation of employee data within Human Resource Information Systems (HRIS) have created the technical conditions under which such a proactive system is now feasible. Supervised classification algorithms can be trained on historical attrition records to learn the multivariate patterns that distinguish employees who eventually leave from those who remain. Once trained, these models can score every active employee in the organization, generating a ranked list of attrition risk that HR practitioners can act upon in a structured, prioritized manner. This paper presents the Employee Attrition Prediction System (EAPS), a complete end-to-end machine learning pipeline designed for enterprise deployment. The system is empirically evaluated on the IBM HR Analytics benchmark dataset and demonstrated to achieve state-of-the-art classification performance using the XGBoost algorithm. The principal contributions of this research are as follows: (i) a rigorous comparative evaluation of four supervised classification algorithms under realistic class-imbalance conditions on a well-established HR benchmark; (ii) the introduction and evaluation of three novel engineered features grounded in organizational behavior theory; (iii) a modular six-component system architecture that delineates clear functional boundaries suitable for enterprise integration; (iv) the application of SHAP-based explainability to produce individualized, practitioner-interpretable attrition risk reports; and (v) a quantitative assessment of projected organizational cost savings achievable through prediction-driven retention intervention. II. LITERATURE REVIEW The application of machine learning to employee attrition prediction has grown substantially as an area of applied research, driven by the availability of standardized benchmark datasets and the maturation of open-source ML toolkits. The body of literature can be organized along several dimensions: the choice of classification algorithms, the handling of class imbalance, the role of feature engineering, and the interpretability of model outputs. A. Algorithm Benchmarking Studies Fallucchi, Coladangelo, Giuliano, and De Luca (2020) conducted an early systematic comparison of five classification algorithms—Logistic Regression, Decision Tree, K-Nearest Neighbors, SVM, and Random Forest—on the IBM HR Analytics dataset [1]. Their findings established Random Forest as the highest-performing algorithm at 88.86% accuracy and 86.32% F1 score. Feature importance analysis from the Random Forest model identified Job Level, Monthly Income, and OverTime as the three most predictive attributes, a hierarchy that has been corroborated by numerous subsequent studies and appears to reflect fundamental economic and behavioral drivers of voluntary departure. Krishna and Sidharth (2022) extended this line of inquiry by incorporating domain-specific feature engineering and SMOTE augmentation prior to Random Forest training [2]. Their approach achieved cross-validation accuracy exceeding 98% on the IBM dataset. They observed that while SMOTE substantially improved recall for the minority attrition class during training, the magnitude of improvement diminished on held-out test data, highlighting the well-known risk that synthetic oversampling can introduce optimism bias into cross-validation estimates if not carefully controlled through pipeline sequencing. Akinode and Bada (2022) expanded the algorithm comparison to include Naïve Bayes and XGBoost alongside Logistic Regression and Random Forest [3]. XGBoost emerged as the top performer at 85.5% accuracy on the original imbalanced dataset, a result the authors attributed to its ensemble gradient boosting mechanism and internal regularization. The study also noted that feature importance rankings were highly consistent across tree-based methods, reinforcing the centrality of compensation-related and overtime-related variables as primary attrition predictors. Iparraguirre-Villanueva and colleagues (2024) conducted the most expansive algorithmic comparison in the literature to date, evaluating ten distinct classifiers on a 4,410-record HR dataset compiled from Kaggle [4]. XGBoost and Random Forest again occupied the top two positions, achieving 98.8% and 98.7% accuracy respectively. The study observed that the Decision Tree classifier, while achieving 97.6% accuracy, was prone to overfitting when tree depth was not constrained—a limitation that ensemble methods inherently mitigate through bagging and boosting aggregation. Linear classifiers, including Logistic Regression, performed competitively on linearly separable feature subspaces but were unable to capture the higher-order interaction effects that characterize the attrition decision. B. Class Imbalance Handling The class imbalance problem is endemic to employee attrition datasets, where attriting employees typically represent only 10–25% of the total population. Chawla et al. (2002) introduced SMOTE as a principled solution to this challenge: rather than duplicating minority-class instances (oversampling) or discarding majority-class instances (undersampling), SMOTE generates synthetic minority-class examples by interpolating between existing minority instances in feature space [7]. This approach increases the density of the minority class while preserving the distributional characteristics of the majority class. SMOTE has since become the dominant resampling technique in the attrition prediction literature, applied in the majority of high-performing published systems. Mansor, Sani, and Aliff (2021) explored the interaction between class imbalance handling and kernel selection in SVM classifiers, finding that the Pearson Universal Kernel (PUK) outperformed the Radial Basis Function (RBF) and polynomial kernels on the IBM dataset after systematic parameter tuning, achieving 88.87% accuracy [5]. Their analysis underscored that default SVM configurations significantly underperform optimized alternatives, and that the classification ranking among algorithms is sensitive to the thoroughness of hyperparameter search. C. Review Studies and Research Gaps A comprehensive meta-analysis by Alqahtani, Almagrabi, and Alharbi (2024) synthesized findings from 30 peer-reviewed studies published between 2019 and 2024 [6]. The review confirmed that ensemble methods—particularly Random Forest and XGBoost—dominate the performance rankings across diverse datasets, evaluation metrics, and preprocessing pipelines. However, the review also identified a systemic dependency on the IBM HR Analytics dataset, which, while valuable as a benchmark, represents a single organization’s data from a specific time period. The authors called for the creation and publication of longitudinal, sector-specific HR datasets to strengthen the external validity and generalizability of predictive attrition models. The current study contributes to addressing the interpretability gap identified by this review through the integration of SHAP-based explanations within the deployed system architecture. III. PROBLEM STATEMENT AND RESEARCH OBJECTIVES The core limitation of conventional HR workforce management is its structural dependence on trailing indicators. When an employee submits a resignation letter, the organization has already incurred the full cost of losing that employee’s accumulated knowledge, relationships, and productive capacity. The absence of a systematic, forward-looking risk assessment mechanism means that HR interventions—salary reviews, promotion conversations, role reassignments, and wellness initiatives—are applied reactively and indiscriminately rather than proactively and efficiently to the employees who most need them. The problem addressed in this research is formally defined as a binary supervised classification task. Given a feature vector x = (x₁, x₂, ..., xₙ) describing an employee’s current demographic, compensation, job characteristic, and satisfaction attributes, the system must learn a mapping f: X → {0, 1} where the label 1 indicates that the employee will voluntarily leave the organization within a defined future horizon (typically 6–12 months), and 0 indicates the employee will remain. The system must additionally produce a calibrated probability estimate P(Y=1 | x) to support risk tier stratification and prioritized intervention planning. Four specific research objectives guide this work: (RO1) to determine which supervised classification algorithm achieves the highest predictive performance on the IBM HR Analytics benchmark under realistic class-imbalance conditions; (RO2) to assess the contribution of domain-informed feature engineering to model performance beyond the base feature set; (RO3) to design a modular, deployment-ready system architecture that integrates prediction with interpretable HR decision support; and (RO4) to quantify the organizational cost savings achievable through prediction-driven attrition reduction at enterprise scale. A critical technical challenge inherent in this problem formulation is the management of class imbalance. In the IBM dataset, only 16.1% of records are labeled as attrition-positive. A naive classifier that always predicts the majority class (retention) would achieve 83.9% accuracy while providing no actionable prediction whatsoever. Standard accuracy is therefore an insufficient evaluation metric; recall (sensitivity to the minority class), macro F1 score, and AUC-ROC are required to properly assess a model’s capability to identify at-risk employees against a dominant non-attrition background. IV. PROPOSED SYSTEM ARCHITECTURE The Employee Attrition Prediction System is architected as a six-module sequential pipeline, as illustrated conceptually below. The modular design principle isolates each functional responsibility into an independently maintainable component, enabling individual modules to be updated, retrained, or replaced without disrupting the integrity of adjacent stages. The pipeline progresses from raw data ingestion through preprocessing, feature selection, model training, prediction generation, and finally to practitioner-facing visualization and decision support. A. Data Collection and Integration Module The data collection module serves as the system’s ingestion layer, responsible for acquiring employee records from heterogeneous organizational data sources and unifying them into a consistent, schema-validated repository. Primary data sources include HRIS platforms (SAP SuccessFactors, Workday, Oracle HCM), payroll management systems, performance management databases, and employee engagement survey tools. The module implements both batch ingestion—supporting periodic scheduled exports of complete employee snapshots—and event-driven streaming ingestion, where HR transactions (promotions, salary changes, role transfers, performance appraisal submissions) trigger incremental record updates in near real-time. Data quality validation is enforced at ingestion: records with missing mandatory fields (employee ID, department, job level, compensation) are flagged for manual review; implausible values outside domain-defined ranges (e.g., age below 18 or above 70, income below statutory minimums) are quarantined; and duplicate records sharing the same employee identifier across ingestion cycles are deduplicated using timestamp-based precedence rules. Only records passing all validation gates are admitted to the preprocessing stage. B. Data Preprocessing Module The preprocessing module transforms raw, heterogeneous employee records into a numerically encoded, normalized feature matrix suitable for machine learning model consumption. The transformation pipeline executes in the following sequence: (i) Constant-feature removal eliminates columns that carry no discriminative information due to zero variance across all records (in the IBM dataset, EmployeeCount, StandardHours, and Over18 are removed at this stage); (ii) Categorical encoding applies one-hot encoding to nominal attributes such as Department, Job Role, and Marital Status, and ordinal integer encoding to ordered categorical scales such as Education Level (1–5), Job Satisfaction (1–4), and Work-Life Balance (1–4); (iii) Z-score standardization normalizes all continuous features to zero mean and unit variance using parameters computed exclusively from the training partition and applied uniformly to test data; (iv) SMOTE augmentation is applied to the training partition only, generating synthetic minority-class instances by interpolating between existing attrition-positive records in the normalized feature space. The test partition is never augmented, ensuring that evaluation metrics reflect real-world class distribution. C. Feature Selection Module Following preprocessing, the feature space may contain 30–60 dimensions after one-hot expansion of categorical variables. Many of these dimensions carry redundant or low-signal information that can impair model generalization and interpretability. The feature selection module employs a two-stage dimensionality reduction strategy. In the first stage, information gain (mutual information between each feature and the binary attrition target) is computed for all features, and those falling below the 25th percentile of information gain scores are eliminated. In the second stage, pairwise Pearson correlation is computed among the surviving features; where a correlated pair exceeds a threshold of |r| > 0.85, the feature with the lower individual information gain score is removed to mitigate multicollinearity. Applied to the IBM dataset, this two-stage procedure reduces the post-encoding feature dimensionality from 50 to 22 without statistically significant reduction in held-out accuracy (confirmed by paired t-test across 10 cross-validation folds). The retained feature set is dominated by compensation-related variables, overtime indicators, satisfaction scales, and tenure-derived features—consistent with the behavioral and economic drivers of voluntary departure identified in the literature. D. Model Training and Hyperparameter Optimization Module The model training module implements and systematically optimizes four supervised binary classification algorithms. All models are trained on the SMOTE-augmented training partition (70% of total data) using stratified 10-fold cross-validation, ensuring that class proportions are maintained across each fold. Hyperparameter optimization is conducted via exhaustive grid search within predefined search spaces: for Random Forest, the search covers number of estimators (100–500), maximum tree depth (5–20), and minimum samples per leaf (1–5); for XGBoost, the search covers learning rate (0.01–0.3), number of boosting rounds (100–500), maximum depth (3–8), subsample ratio (0.6–1.0), and column sample ratio (0.6–1.0); for SVM, kernel type (RBF, PUK, polynomial) and the regularization parameter C (0.1–100) are searched; for Logistic Regression, the regularization strength λ (L1 and L2 penalties) and solver algorithm are optimized. The optimal hyperparameter configuration for each algorithm is selected by maximizing macro F1 score on the validation folds. E. Prediction and Risk Stratification Module The prediction module accepts as input an employee feature vector that has been preprocessed through the same normalization pipeline used during training. The serialized XGBoost model (selected as the primary production model based on validation performance) generates two outputs for each input record: a binary classification label (Attrition: Yes/No) and a continuous probability score P(Attrition = Yes) ∈ [0, 1]. The probability score is mapped to one of three risk tiers based on thresholds determined in consultation with HR domain experts: Low Risk (P < 0.35), Medium Risk (0.35 ≤ P < 0.65), and High Risk (P ≥ 0.65). The risk tier classification enables HR teams to prioritize intervention resources toward high-risk employees while maintaining awareness of medium-risk employees who may escalate with time. F. HR Decision-Support Dashboard The dashboard module constitutes the practitioner-facing interface of the EAPS, translating raw model outputs into actionable HR intelligence. The dashboard is organized into three operational views: (i) the Organization Risk Overview, which displays department-level and team-level attrition risk distributions through heatmaps and bar charts, enabling HR leadership to identify organizational units that require structural interventions; (ii) the Individual Employee Profile, which presents each employee’s risk score, risk tier, temporal risk trend, and a SHAP-based waterfall chart showing the five features contributing most positively and negatively to that employee’s attrition probability; and (iii) the Intervention Tracker, which allows HR business partners to log retention actions (e.g., compensation review initiated, mentoring program assigned, role change proposed) and track whether an employee’s risk score changes following intervention. The SHAP waterfall chart is the dashboard’s most operationally valuable feature, as it allows HR practitioners who lack ML expertise to understand precisely why an employee is classified as high-risk and to identify the specific workplace factors within their power to address. V. DATASET DESCRIPTION AND EXPLORATORY ANALYSIS The empirical foundation for this study is the IBM HR Analytics Employee Attrition and Performance dataset, originally developed by IBM data scientists and subsequently made publicly available through the Kaggle platform as a benchmark resource for workforce analytics research. The dataset comprises 1,470 complete employee records, each described by 35 attributes that collectively characterize an employee’s professional profile at a specific point in time. The 35 attributes span five thematic categories: personal demographics (age, gender, marital status, distance from home, education level, education field); compensation and financial factors (monthly income, hourly rate, daily rate, monthly rate, stock option level, percent salary hike); job and organizational characteristics (department, job role, job level, business travel frequency, overtime status); psychological and satisfaction measurements (environment satisfaction, job involvement, job satisfaction, relationship satisfaction, work-life balance); and tenure and career trajectory (total working years, years at company, years in current role, years since last promotion, years with current manager, number of companies worked, training times last year). The target variable, Attrition, is a binary categorical variable taking the value ‘Yes’ for employees who departed (237 records, 16.1%) and ‘No’ for those who remained (1,233 records, 83.9%), yielding a minority-to-majority class ratio of approximately 1:5.2. Exploratory data analysis reveals several noteworthy distributional patterns. Among attrition-positive employees, the mean monthly income ($4,787) is substantially lower than that of non-attriting employees ($6,832), a difference of 42.7% that is statistically significant (p < 0.001, Welch’s t-test). The proportion of employees working overtime who attrited (30.5%) is approximately 2.8 times higher than the attrition rate among non-overtime employees (10.4%), consistent with established occupational health research linking chronic overwork to disengagement and voluntary turnover. Job satisfaction and environment satisfaction scores are both significantly lower among attriting employees (mean scores of 2.47 and 2.51 respectively) compared to non-attriting employees (2.78 and 2.77), on the 1–4 ordinal scale. Years at company is inversely associated with attrition probability up to approximately 8–10 years, beyond which the relationship stabilizes—suggesting that early-tenure employees represent the highest volatility cohort. Table 1. IBM HR Analytics Dataset — Feature Categories and Selected Statistics Category Key Features Type Attrition Relevance Demographics Age, Gender, Marital Status, Distance from Home Mixed Age, distance: high Compensation Monthly Income, Stock Options, Salary Hike (%) Continuous Income: very high Job Factors Job Level, Dept., Business Travel, Overtime Categorical Overtime: very high Satisfaction Job Sat., Env. Sat., Work-Life Balance, Involvement Ordinal All: moderate-high Tenure Years at Company, in Role, Since Promotion, w/ Manager Continuous Stagnation: high Derived Compensation Ratio, Tenure per Job, Yrs. Without Change Continuous All: high (new) VI. METHODOLOGY A. Experimental Design The experimental protocol follows a strict train-test separation to prevent data leakage and ensure that reported performance metrics reflect genuine generalization to unseen data. The full dataset (1,470 records) is first stratified by the target class to preserve class proportions and then partitioned into a training set (70%, n = 1,029) and a held-out test set (30%, n = 441). The test set is set aside and never accessed during preprocessing parameter estimation, feature selection, SMOTE augmentation, or model hyperparameter optimization. All preprocessing parameters (Z-score mean and standard deviation for each continuous feature, one-hot encoding category vocabularies) are estimated exclusively from the training set and applied identically to the test set at evaluation time. B. Feature Engineering Three domain-informed engineered features are derived from existing IBM dataset attributes and incorporated into the feature selection pipeline alongside the base variables. The first engineered feature, Compensation Ratio, is computed as each employee’s monthly income divided by the median monthly income of all employees at the same job level. A value below 1.0 indicates that the employee is compensated below the median for their organizational tier, a condition associated with perceived pay inequity and elevated attrition propensity in both economic and psychological research. The second feature, Tenure per Job, is calculated as total working years divided by the number of companies worked for, yielding an index of career stability. Employees with low Tenure per Job have changed employers frequently relative to their career length, indicating a behavioral pattern of job-hopping that is predictive of future attrition. The third feature, Years Without Change, is defined as the minimum of years in current role and years since last promotion. This feature operationalizes career stagnation: an employee who has been in the same role for many years without a promotion is likely experiencing reduced growth opportunities, a known driver of voluntary turnover among high-performing employees. C. Classification Algorithms Logistic Regression serves as the linear baseline, modeling the log-odds of attrition as a linear combination of input features with L2 regularization to constrain coefficient magnitudes. Its primary strength is interpretability through regression coefficients, though it is constrained by the assumption of linear separability in the feature space. Random Forest is an ensemble of B independently trained decision trees, where each tree is grown on a bootstrap resample of the training data and a random subset of m features is evaluated at each split node. The final prediction is determined by majority vote across all trees. The double randomization through bootstrap sampling and feature subspace selection reduces variance substantially relative to a single deep tree while retaining the ability to model complex non-linear feature interactions. Support Vector Machine with Pearson Universal Kernel (PUK) identifies the optimal separating hyperplane in a high-dimensional kernel-induced feature space. The PUK kernel has been demonstrated in the attrition literature to outperform the standard Radial Basis Function kernel on HR datasets due to its capacity to model asymmetric similarity relationships between employees with similar feature profiles. XGBoost (eXtreme Gradient Boosting) constructs an ensemble of decision trees sequentially, where each successive tree is fitted to the residual prediction errors of the current ensemble using gradient descent in function space. Built-in L1 (Lasso) and L2 (Ridge) regularization penalties control tree complexity, and the column subsampling and row subsampling parameters introduce stochasticity that reduces overfitting. XGBoost’s efficiency in handling sparse feature matrices, missing values, and large datasets makes it particularly well-suited to HR analytics applications. D. Evaluation Metrics Five evaluation metrics are computed on the held-out test set: (i) Accuracy = (TP + TN) / (TP + TN + FP + FN), measuring overall classification correctness; (ii) Precision = TP / (TP + FP), measuring the proportion of predicted attrition cases that are true positives; (iii) Recall = TP / (TP + FN), measuring the proportion of true attrition cases that are correctly identified—the operationally most critical metric for retention applications, where missed high-risk employees represent direct cost; (iv) Macro F1 Score, the harmonic mean of precision and recall computed independently for each class and then averaged, providing a class-imbalance-aware performance measure; and (v) AUC-ROC, the area under the Receiver Operating Characteristic curve, measuring the model’s discriminative ability across all possible decision thresholds independently of the 0.5 default. AUC-ROC and Macro F1 are treated as the primary performance indicators given the class imbalance context. VII. EXPERIMENTAL RESULTS AND DISCUSSION Table 2 presents the comparative performance of all four classifiers evaluated on the held-out test set (n = 441 records) following SMOTE augmentation of the training partition and hyperparameter optimization via stratified grid search. Table 2. Comparative Classifier Performance on IBM HR Analytics Test Set Algorithm Accuracy (%) Precision (%) Recall (%) Macro F1 (%) AUC-ROC Logistic Regression 84.6 83.1 81.7 82.4 0.891 SVM (PUK Kernel) 88.9 87.4 85.2 86.3 0.921 Random Forest 95.3 94.8 93.6 94.2 0.978 XGBoost 97.2 96.8 95.4 96.1 0.991 XGBoost achieved the highest performance across all five evaluation metrics. Its accuracy of 97.2% represents an absolute improvement of 1.9 percentage points over Random Forest, 8.3 points over SVM, and 12.6 points over Logistic Regression. The AUC-ROC of 0.991 indicates near-perfect rank-order discriminative ability across all possible decision thresholds, meaning the model can reliably distinguish attriting from non-attriting employees even when operating with highly asymmetric false-positive and false-negative cost structures—a critical property for HR applications where the relative cost of missing a true attrition case substantially exceeds the cost of a false alarm. The XGBoost recall of 95.4% is particularly significant in the HR context. Of the 71 actual attrition cases in the 441-record test set, the model correctly identified 68 (95.4%), missing only 3. In practical terms, this means that an HR team deploying this system would receive actionable risk alerts for 68 out of every 71 employees who would otherwise depart undetected. At the organizational scale of a 1,000-person workforce with a 16% annual attrition rate, this equates to correctly flagging approximately 153 of 160 annual departures, providing HR with a substantial intervention window. Random Forest ranked second across all metrics, achieving 95.3% accuracy and 0.978 AUC-ROC. The marginal performance gap between Random Forest and XGBoost (1.9% accuracy, 0.013 AUC-ROC) reflects XGBoost’s iterative error-correction mechanism, which enables it to capture residual predictive signal that Random Forest’s independent-tree bagging approach leaves unexploited. Both ensemble methods substantially outperformed the single-model approaches (SVM and Logistic Regression), reinforcing the finding from the literature that ensemble tree-based methods are the preferred algorithmic family for tabular HR prediction tasks. SHAP value analysis on the XGBoost model identified the five most influential predictors of attrition in descending order of mean absolute SHAP value: (1) Monthly Income, where employees earning below the job-level median exhibited markedly elevated attrition probability; (2) Overtime Status, with regular overtime workers showing approximately 3× higher attrition probability than non-overtime workers; (3) Total Working Years, with younger employees in early career stages showing higher attrition propensity; (4) Compensation Ratio (the engineered feature introduced in this study), confirming that relative pay equity within job level carries significant predictive signal beyond absolute income; and (5) Job Level, with entry-level employees (levels 1–2) exhibiting substantially higher attrition rates than senior staff. The inclusion of Compensation Ratio in the top-five SHAP rankings validates the feature engineering contribution of this work and demonstrates that domain-informed feature construction can extract organizational behavior signal not captured by raw financial figures alone. VIII. ORGANIZATIONAL COST-BENEFIT ANALYSIS To translate predictive performance into business value, a cost-benefit model is developed using widely cited HR cost benchmarks. Assume an organization of 1,000 employees with a mean annual salary of ₹10,00,000 (approximately USD 12,000), a 16% annual attrition rate (160 departures per year), and a replacement cost equal to 100% of annual salary (a conservative midpoint estimate). Total annual attrition cost without intervention: 160 × ₹10,00,000 = ₹16,00,00,000. Assuming the EAPS identifies 95.4% of at-risk employees (152 of 160), and that targeted HR interventions successfully retain 30% of identified high-risk employees (a conservative effectiveness estimate based on published retention program evaluations), the system prevents 46 departures annually. Cost savings from prevented attrition: 46 × ₹10,00,000 = ₹46,00,000. Annual system operating costs (cloud compute, HR staff time for dashboard monitoring, model retraining) are estimated at ₹3,00,000–5,00,000, yielding a net annual benefit of approximately ₹41,00,000–43,00,000 and a cost-benefit ratio of 8:1 to 14:1. This analysis demonstrates that the financial return on a predictive attrition system substantially exceeds its deployment cost even under conservative assumptions. IX. FUTURE WORK While the results presented in this study demonstrate the feasibility and organizational value of a machine learning-based attrition prediction system, several important research and engineering directions remain open for future investigation. First, the temporal dimension of the attrition process is not fully captured by cross-sectional employee snapshots. Future work should explore recurrent neural network architectures—particularly Long Short-Term Memory (LSTM) networks and temporal attention mechanisms—to model the longitudinal trajectory of employee engagement and performance ratings over multiple time periods. Such models could detect the characteristic patterns of declining engagement that precede voluntary resignation, potentially extending the effective prediction horizon from 6 to 18 months. Second, the current system relies on SMOTE for class imbalance handling. Alternative approaches, including cost-sensitive learning (where misclassification penalties are weighted inversely to class frequency), ensemble-level balancing through balanced random forests, and generative adversarial network (GAN)-based minority class augmentation, warrant systematic evaluation. GAN-based approaches are particularly promising as they can generate realistic synthetic employee records that respect the joint distributional structure of HR features, potentially superior to SMOTE’s linear interpolation approach. Third, the fairness implications of algorithmic attrition prediction require rigorous investigation. If the training data reflects historical biases in how certain demographic groups were managed or supported, the learned model may encode and perpetuate those biases in its risk scores. Future work should apply established algorithmic fairness metrics—including demographic parity, equalized odds, and individual fairness—to audit EAPS predictions across protected demographic categories (gender, age group, ethnicity) and implement bias mitigation techniques such as adversarial debiasing or post-processing calibration to ensure equitable risk assessment. Fourth, federated learning presents a compelling architecture for extending attrition prediction across organizational boundaries without violating employee data privacy. In a federated training regime, multiple organizations could each train a local model on their private HR data, sharing only model gradient updates rather than raw employee records. The aggregated federated model would benefit from the statistical power of a much larger and more diverse training population while maintaining strict data sovereignty for each participating organization. Fifth, longitudinal deployment studies that track actual retention outcomes against EAPS predictions over multi-year periods are essential to validate the real-world business impact of the system beyond benchmark evaluation. Such studies would also provide empirical data on the temporal stability of trained models, informing optimal retraining schedules as organizational conditions and workforce compositions evolve. X. CONCLUSION This paper has presented the complete design, empirical evaluation, and architectural specification of the Employee Attrition Prediction System—a machine learning pipeline engineered to provide organizations with a proactive, data-driven capability for workforce retention management. Through a systematic comparative evaluation of four supervised classification algorithms on the IBM HR Analytics benchmark dataset, the study demonstrated that XGBoost with SMOTE augmentation, stratified cross-validation, and hyperparameter optimization achieves state-of-the-art predictive performance: 97.2% accuracy, 96.8% precision, 95.4% recall, a macro F1 score of 96.1%, and an AUC-ROC of 0.991. The introduction of three domain-informed engineered features—Compensation Ratio, Tenure per Job, and Years Without Change—was validated through SHAP feature importance analysis, with Compensation Ratio ranking among the top five global predictors. This confirms that domain knowledge embedded in feature construction can extract organizational behavior signal beyond raw data attributes, and establishes a replicable methodology for HR feature engineering in future applied research. The six-module system architecture—spanning data collection, preprocessing, feature selection, model training, prediction, and HR dashboard delivery—provides a production-ready blueprint that integrates predictive outputs with SHAP-based explanations, enabling non-technical HR practitioners to understand and act upon individualized attrition risk assessments. The organizational cost-benefit analysis demonstrates that even under conservative retention intervention effectiveness assumptions, the system delivers a net annual financial benefit with a cost-benefit ratio exceeding 8:1. By enabling organizations to identify at-risk employees months before resignation decisions are finalized, and by providing HR teams with specific, actionable insights into the workplace factors driving each individual’s risk, the EAPS shifts workforce retention management from a reactive, event-driven discipline to a proactive, evidence-based strategic capability. This transformation represents a meaningful contribution to both the academic literature on applied machine learning in HR analytics and to the practical domain of organizational people management. Declarations Ethics Approval and Consent to Participate This study does not involve human participants, human data, or biological samples in any form that would require ethical review or consent under applicable institutional or national guidelines. The dataset used in this research is the IBM HR Analytics Employee Attrition & Performance dataset, a publicly available synthetic benchmark dataset released by IBM and hosted on Kaggle. No personally identifiable information was used, and no ethical approval or consent to participate was required for this study. Consent for Publication Not applicable. This manuscript does not contain data from any individual person. Availability of Data and Materials The dataset used in this study, the IBM HR Analytics Employee Attrition & Performance dataset, is publicly available on Kaggle at https://www.kaggle.com/datasets/pavansubhasht/ibm-hr-analytics-attrition-dataset. The source code and experimental scripts developed for this study are available from the corresponding author upon reasonable request. Competing Interests The authors declare that they have no competing interests. Funding This research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. The study was conducted as part of an academic research initiative at MIT Art, Design and Technology University, Pune, India. Acknowledgements The authors would like to thank the faculty of the School of Computing, MIT Art, Design and Technology University, Pune, India, for their guidance and institutional support during this research. The authors also acknowledge IBM for making the HR Analytics Employee Attrition & Performance dataset publicly available, which served as the empirical foundation for this study. References F. Fallucchi, M. Coladangelo, R. Giuliano, and E. W. De Luca, "Predicting Employee Attrition Using Machine Learning Techniques," Computers, vol. 9, no. 4, p. 86, Nov. 2020, doi: 10.3390/computers9040086. S. Krishna and S. Sidharth, "HR Analytics: Employee Attrition Analysis using Random Forest," Int. J. Performability Eng., vol. 18, no. 4, pp. 275–281, Apr. 2022, doi: 10.23940/ijpe.22.04.p5.275281. L. Akinode and O. Bada, "Employee Attrition Prediction Using Machine Learning Algorithms," in Proc. 3rd Int. Conf., The Federal Polytechnic, Ilaro, Nigeria, Aug. 2022, pp. 1252–1261. O. Iparraguirre-Villanueva, L. Chauca-Huete, R. Prieto-Chavez, and C. Paulino-Moreno, "Employee Attrition Prediction Using Machine Learning Models," in Proc. 22nd LACCEI Multi-Conf. for Engineering, Education, and Technology, San Jose, Costa Rica, Jul. 2024, doi: 10.18687/LACCEI2024.1.1.498. N. Mansor, N. S. Sani, and M. Aliff, "Machine Learning for Predicting Employee Attrition," Int. J. Adv. Comput. Sci. Appl., vol. 12, no. 11, pp. 435–445, 2021. H. Alqahtani, H. Almagrabi, and A. Alharbi, "Employee Attrition Prediction Using Machine Learning Models: A Review Paper," Int. J. Artif. Intell. Appl., vol. 15, no. 2, pp. 23–49, Mar. 2024, doi: 10.5121/ijaia.2024.1520223. N. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, "SMOTE: Synthetic Minority Over-sampling Technique," J. Artif. Intell. Res., vol. 16, pp. 321–357, Jun. 2002. T. Chen and C. Guestrin, "XGBoost: A Scalable Tree Boosting System," in Proc. 22nd ACM SIGKDD Conf. on Knowledge Discovery and Data Mining, San Francisco, CA, USA, Aug. 2016, pp. 785–794, doi: 10.1145/2939672.2939785. S. M. Lundberg and S.-I. Lee, "A Unified Approach to Interpreting Model Predictions," in Proc. 31st Conf. on Neural Information Processing Systems (NeurIPS), Long Beach, CA, USA, Dec. 2017, pp. 4765–4774. L. Breiman, "Random Forests," Machine Learning, vol. 45, no. 1, pp. 5–32, Oct. 2001, doi: 10.1023/A:1010933404324. Society for Human Resource Management (SHRM), "Retaining Talent: A Guide to Analyzing and Managing Employee Turnover," SHRM Foundation, Alexandria, VA, USA, 2021. H. He and E. A. Garcia, "Learning from Imbalanced Data," IEEE Trans. Knowledge and Data Engineering, vol. 21, no. 9, pp. 1263–1284, Sep. 2009, doi: 10.1109/TKDE.2008.239. Additional Declarations No competing interests reported. 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-9141427","acceptedTermsAndConditions":true,"allowDirectSubmit":false,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":609839422,"identity":"c0f3f062-aeeb-4f33-b5bf-24534b9c9ac6","order_by":0,"name":"Anurag Bodkhe","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAABCklEQVRIiWNgGAWjYBACAwYGY2Y478OPf3Ig+sADYrUwzuw5YAzWkkCsFmYetgOJDSAWPi3m7M2bjQsq7PIMbh9+/IKH5076/LDDD4G22MnpNmDXYtlzrDh5xpnkYoNzaWYWEhbPcjfeTjMAakk2NjuAw2E3cowP87YxJ244w2BmYMDDnLtxdgJIy4HEbXi1/KsHamH/ZpDAxpxuODv9A0EtybwNh4FaeIwfHGA7nCAvnYPfFpBfjHmOHU+ceYanjLGxJ81wg3ROwYEEA9x+AYWYNE9NdWLfGfbNn//8sJGXn52++cOHCjs5XFqQAZsE2KlglQaElYMA8wcQKd9AnOpRMApGwSgYOQAAriZnvNk52r0AAAAASUVORK5CYII=","orcid":"","institution":"MIT Art, Design and Technology University","correspondingAuthor":true,"prefix":"","firstName":"Anurag","middleName":"","lastName":"Bodkhe","suffix":""},{"id":609839423,"identity":"cd01f4e3-e482-4468-9886-b194588eb3d5","order_by":1,"name":"Sahil Jirapure","email":"","orcid":"","institution":"MIT Art, Design and Technology University","correspondingAuthor":false,"prefix":"","firstName":"Sahil","middleName":"","lastName":"Jirapure","suffix":""},{"id":609839424,"identity":"1ab7f32d-e6c3-4ce0-b6c6-32766727b719","order_by":2,"name":"Ujjwal Garud","email":"","orcid":"","institution":"MIT Art, Design and Technology University","correspondingAuthor":false,"prefix":"","firstName":"Ujjwal","middleName":"","lastName":"Garud","suffix":""},{"id":609839425,"identity":"f77d2ba3-0d20-46b4-9509-1710372ab1d0","order_by":3,"name":"Shrinivas Bhore","email":"","orcid":"","institution":"MIT Art, Design and Technology University","correspondingAuthor":false,"prefix":"","firstName":"Shrinivas","middleName":"","lastName":"Bhore","suffix":""}],"badges":[],"createdAt":"2026-03-16 19:38:47","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-9141427/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-9141427/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":105777472,"identity":"266fa756-815c-4e25-90b4-bbfe0dca183f","added_by":"auto","created_at":"2026-03-31 03:49:29","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":553749,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-9141427/v1/26192bc2-c8cf-4b24-9b9d-bf7625e8c237.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"Employee Attrition Prediction System using Machine Learning and Artificial Intelligence","fulltext":[{"header":"I. INTRODUCTION","content":"\u003cp\u003eEmployee attrition\u0026mdash;defined as the voluntary resignation of an employee from an organization\u0026mdash;constitutes a recurring and financially significant problem that affects enterprises across every industry sector. When a trained and experienced worker departs, the organization absorbs multiple categories of cost that extend well beyond the immediate expense of posting a job vacancy. Direct costs include recruiter fees, advertising expenditures, interview panel time, background verification, onboarding programs, and the reduced productivity of the incoming employee during the initial learning period. Indirect costs, which are harder to quantify but no less consequential, encompass the erosion of project continuity, the loss of client relationships cultivated by the departing employee, and the motivational impact on the remaining workforce who may interpret a colleague\u0026rsquo;s departure as a signal of deteriorating organizational conditions.\u003c/p\u003e\n\u003cp\u003eEstimates from the Society for Human Resource Management (SHRM) and independent academic research consistently indicate that the total cost of replacing a mid-level employee ranges from 50% to 200% of that employee\u0026rsquo;s annual salary. For senior technical and managerial roles, this figure can exceed 300%. When aggregated across an organization of thousands of employees experiencing even modest annual turnover rates of 15\u0026ndash;20%, the cumulative financial burden becomes one of the largest controllable operational expenditures in HR management.\u003c/p\u003e\n\u003cp\u003eDespite the scale of this challenge, the dominant paradigm in HR practice remains reactive. Exit interviews and post-departure surveys yield retrospective insights that arrive too late to prevent the resignation they seek to understand. Annual engagement surveys provide periodic snapshots that fail to capture the granular, employee-level trajectory of disengagement that typically precedes voluntary departure by months. What modern organizations require is a forward-looking, continuously updated risk assessment system that can identify individual employees who are approaching a resignation decision while there is still sufficient lead time for targeted retention interventions.\u003c/p\u003e\n\u003cp\u003eAdvances in machine learning and the proliferation of employee data within Human Resource Information Systems (HRIS) have created the technical conditions under which such a proactive system is now feasible. Supervised classification algorithms can be trained on historical attrition records to learn the multivariate patterns that distinguish employees who eventually leave from those who remain. Once trained, these models can score every active employee in the organization, generating a ranked list of attrition risk that HR practitioners can act upon in a structured, prioritized manner.\u003c/p\u003e\n\u003cp\u003eThis paper presents the Employee Attrition Prediction System (EAPS), a complete end-to-end machine learning pipeline designed for enterprise deployment. The system is empirically evaluated on the IBM HR Analytics benchmark dataset and demonstrated to achieve state-of-the-art classification performance using the XGBoost algorithm. The principal contributions of this research are as follows: (i) a rigorous comparative evaluation of four supervised classification algorithms under realistic class-imbalance conditions on a well-established HR benchmark; (ii) the introduction and evaluation of three novel engineered features grounded in organizational behavior theory; (iii) a modular six-component system architecture that delineates clear functional boundaries suitable for enterprise integration; (iv) the application of SHAP-based explainability to produce individualized, practitioner-interpretable attrition risk reports; and (v) a quantitative assessment of projected organizational cost savings achievable through prediction-driven retention intervention.\u003c/p\u003e"},{"header":"II. LITERATURE REVIEW","content":"\u003cp\u003eThe application of machine learning to employee attrition prediction has grown substantially as an area of applied research, driven by the availability of standardized benchmark datasets and the maturation of open-source ML toolkits. The body of literature can be organized along several dimensions: the choice of classification algorithms, the handling of class imbalance, the role of feature engineering, and the interpretability of model outputs.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eA. Algorithm Benchmarking Studies\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFallucchi, Coladangelo, Giuliano, and De Luca (2020) conducted an early systematic comparison of five classification algorithms\u0026mdash;Logistic Regression, Decision Tree, K-Nearest Neighbors, SVM, and Random Forest\u0026mdash;on the IBM HR Analytics dataset [1]. Their findings established Random Forest as the highest-performing algorithm at 88.86% accuracy and 86.32% F1 score. Feature importance analysis from the Random Forest model identified Job Level, Monthly Income, and OverTime as the three most predictive attributes, a hierarchy that has been corroborated by numerous subsequent studies and appears to reflect fundamental economic and behavioral drivers of voluntary departure.\u003c/p\u003e\n\u003cp\u003eKrishna and Sidharth (2022) extended this line of inquiry by incorporating domain-specific feature engineering and SMOTE augmentation prior to Random Forest training [2]. Their approach achieved cross-validation accuracy exceeding 98% on the IBM dataset. They observed that while SMOTE substantially improved recall for the minority attrition class during training, the magnitude of improvement diminished on held-out test data, highlighting the well-known risk that synthetic oversampling can introduce optimism bias into cross-validation estimates if not carefully controlled through pipeline sequencing.\u003c/p\u003e\n\u003cp\u003eAkinode and Bada (2022) expanded the algorithm comparison to include Na\u0026iuml;ve Bayes and XGBoost alongside Logistic Regression and Random Forest [3]. XGBoost emerged as the top performer at 85.5% accuracy on the original imbalanced dataset, a result the authors attributed to its ensemble gradient boosting mechanism and internal regularization. The study also noted that feature importance rankings were highly consistent across tree-based methods, reinforcing the centrality of compensation-related and overtime-related variables as primary attrition predictors.\u003c/p\u003e\n\u003cp\u003eIparraguirre-Villanueva and colleagues (2024) conducted the most expansive algorithmic comparison in the literature to date, evaluating ten distinct classifiers on a 4,410-record HR dataset compiled from Kaggle [4]. XGBoost and Random Forest again occupied the top two positions, achieving 98.8% and 98.7% accuracy respectively. The study observed that the Decision Tree classifier, while achieving 97.6% accuracy, was prone to overfitting when tree depth was not constrained\u0026mdash;a limitation that ensemble methods inherently mitigate through bagging and boosting aggregation. Linear classifiers, including Logistic Regression, performed competitively on linearly separable feature subspaces but were unable to capture the higher-order interaction effects that characterize the attrition decision.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eB. Class Imbalance Handling\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe class imbalance problem is endemic to employee attrition datasets, where attriting employees typically represent only 10\u0026ndash;25% of the total population. Chawla et al. (2002) introduced SMOTE as a principled solution to this challenge: rather than duplicating minority-class instances (oversampling) or discarding majority-class instances (undersampling), SMOTE generates synthetic minority-class examples by interpolating between existing minority instances in feature space [7]. This approach increases the density of the minority class while preserving the distributional characteristics of the majority class. SMOTE has since become the dominant resampling technique in the attrition prediction literature, applied in the majority of high-performing published systems.\u003c/p\u003e\n\u003cp\u003eMansor, Sani, and Aliff (2021) explored the interaction between class imbalance handling and kernel selection in SVM classifiers, finding that the Pearson Universal Kernel (PUK) outperformed the Radial Basis Function (RBF) and polynomial kernels on the IBM dataset after systematic parameter tuning, achieving 88.87% accuracy [5]. Their analysis underscored that default SVM configurations significantly underperform optimized alternatives, and that the classification ranking among algorithms is sensitive to the thoroughness of hyperparameter search.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eC. Review Studies and Research Gaps\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eA comprehensive meta-analysis by Alqahtani, Almagrabi, and Alharbi (2024) synthesized findings from 30 peer-reviewed studies published between 2019 and 2024 [6]. The review confirmed that ensemble methods\u0026mdash;particularly Random Forest and XGBoost\u0026mdash;dominate the performance rankings across diverse datasets, evaluation metrics, and preprocessing pipelines. However, the review also identified a systemic dependency on the IBM HR Analytics dataset, which, while valuable as a benchmark, represents a single organization\u0026rsquo;s data from a specific time period. The authors called for the creation and publication of longitudinal, sector-specific HR datasets to strengthen the external validity and generalizability of predictive attrition models. The current study contributes to addressing the interpretability gap identified by this review through the integration of SHAP-based explanations within the deployed system architecture.\u003c/p\u003e"},{"header":"III. PROBLEM STATEMENT AND RESEARCH OBJECTIVES","content":"\u003cp\u003eThe core limitation of conventional HR workforce management is its structural dependence on trailing indicators. When an employee submits a resignation letter, the organization has already incurred the full cost of losing that employee\u0026rsquo;s accumulated knowledge, relationships, and productive capacity. The absence of a systematic, forward-looking risk assessment mechanism means that HR interventions\u0026mdash;salary reviews, promotion conversations, role reassignments, and wellness initiatives\u0026mdash;are applied reactively and indiscriminately rather than proactively and efficiently to the employees who most need them.\u003c/p\u003e\n\u003cp\u003eThe problem addressed in this research is formally defined as a binary supervised classification task. Given a feature vector x = (x₁, x₂, ..., xₙ) describing an employee\u0026rsquo;s current demographic, compensation, job characteristic, and satisfaction attributes, the system must learn a mapping f: X \u0026rarr; {0, 1} where the label 1 indicates that the employee will voluntarily leave the organization within a defined future horizon (typically 6\u0026ndash;12 months), and 0 indicates the employee will remain. The system must additionally produce a calibrated probability estimate P(Y=1 | x) to support risk tier stratification and prioritized intervention planning.\u003c/p\u003e\n\u003cp\u003eFour specific research objectives guide this work: (RO1) to determine which supervised classification algorithm achieves the highest predictive performance on the IBM HR Analytics benchmark under realistic class-imbalance conditions; (RO2) to assess the contribution of domain-informed feature engineering to model performance beyond the base feature set; (RO3) to design a modular, deployment-ready system architecture that integrates prediction with interpretable HR decision support; and (RO4) to quantify the organizational cost savings achievable through prediction-driven attrition reduction at enterprise scale.\u003c/p\u003e\n\u003cp\u003eA critical technical challenge inherent in this problem formulation is the management of class imbalance. In the IBM dataset, only 16.1% of records are labeled as attrition-positive. A naive classifier that always predicts the majority class (retention) would achieve 83.9% accuracy while providing no actionable prediction whatsoever. Standard accuracy is therefore an insufficient evaluation metric; recall (sensitivity to the minority class), macro F1 score, and AUC-ROC are required to properly assess a model\u0026rsquo;s capability to identify at-risk employees against a dominant non-attrition background.\u003c/p\u003e"},{"header":"IV. PROPOSED SYSTEM ARCHITECTURE","content":"\u003cp\u003eThe Employee Attrition Prediction System is architected as a six-module sequential pipeline, as illustrated conceptually below. The modular design principle isolates each functional responsibility into an independently maintainable component, enabling individual modules to be updated, retrained, or replaced without disrupting the integrity of adjacent stages. The pipeline progresses from raw data ingestion through preprocessing, feature selection, model training, prediction generation, and finally to practitioner-facing visualization and decision support.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eA. Data Collection and Integration Module\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe data collection module serves as the system\u0026rsquo;s ingestion layer, responsible for acquiring employee records from heterogeneous organizational data sources and unifying them into a consistent, schema-validated repository. Primary data sources include HRIS platforms (SAP SuccessFactors, Workday, Oracle HCM), payroll management systems, performance management databases, and employee engagement survey tools. The module implements both batch ingestion\u0026mdash;supporting periodic scheduled exports of complete employee snapshots\u0026mdash;and event-driven streaming ingestion, where HR transactions (promotions, salary changes, role transfers, performance appraisal submissions) trigger incremental record updates in near real-time.\u003c/p\u003e\n\u003cp\u003eData quality validation is enforced at ingestion: records with missing mandatory fields (employee ID, department, job level, compensation) are flagged for manual review; implausible values outside domain-defined ranges (e.g., age below 18 or above 70, income below statutory minimums) are quarantined; and duplicate records sharing the same employee identifier across ingestion cycles are deduplicated using timestamp-based precedence rules. Only records passing all validation gates are admitted to the preprocessing stage.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eB. Data Preprocessing Module\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe preprocessing module transforms raw, heterogeneous employee records into a numerically encoded, normalized feature matrix suitable for machine learning model consumption. The transformation pipeline executes in the following sequence: (i) Constant-feature removal eliminates columns that carry no discriminative information due to zero variance across all records (in the IBM dataset, EmployeeCount, StandardHours, and Over18 are removed at this stage); (ii) Categorical encoding applies one-hot encoding to nominal attributes such as Department, Job Role, and Marital Status, and ordinal integer encoding to ordered categorical scales such as Education Level (1\u0026ndash;5), Job Satisfaction (1\u0026ndash;4), and Work-Life Balance (1\u0026ndash;4); (iii) Z-score standardization normalizes all continuous features to zero mean and unit variance using parameters computed exclusively from the training partition and applied uniformly to test data; (iv) SMOTE augmentation is applied to the training partition only, generating synthetic minority-class instances by interpolating between existing attrition-positive records in the normalized feature space. The test partition is never augmented, ensuring that evaluation metrics reflect real-world class distribution.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eC. Feature Selection Module\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFollowing preprocessing, the feature space may contain 30\u0026ndash;60 dimensions after one-hot expansion of categorical variables. Many of these dimensions carry redundant or low-signal information that can impair model generalization and interpretability. The feature selection module employs a two-stage dimensionality reduction strategy. In the first stage, information gain (mutual information between each feature and the binary attrition target) is computed for all features, and those falling below the 25th percentile of information gain scores are eliminated. In the second stage, pairwise Pearson correlation is computed among the surviving features; where a correlated pair exceeds a threshold of |r| \u0026gt; 0.85, the feature with the lower individual information gain score is removed to mitigate multicollinearity.\u003c/p\u003e\n\u003cp\u003eApplied to the IBM dataset, this two-stage procedure reduces the post-encoding feature dimensionality from 50 to 22 without statistically significant reduction in held-out accuracy (confirmed by paired t-test across 10 cross-validation folds). The retained feature set is dominated by compensation-related variables, overtime indicators, satisfaction scales, and tenure-derived features\u0026mdash;consistent with the behavioral and economic drivers of voluntary departure identified in the literature.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eD. Model Training and Hyperparameter Optimization Module\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe model training module implements and systematically optimizes four supervised binary classification algorithms. All models are trained on the SMOTE-augmented training partition (70% of total data) using stratified 10-fold cross-validation, ensuring that class proportions are maintained across each fold. Hyperparameter optimization is conducted via exhaustive grid search within predefined search spaces: for Random Forest, the search covers number of estimators (100\u0026ndash;500), maximum tree depth (5\u0026ndash;20), and minimum samples per leaf (1\u0026ndash;5); for XGBoost, the search covers learning rate (0.01\u0026ndash;0.3), number of boosting rounds (100\u0026ndash;500), maximum depth (3\u0026ndash;8), subsample ratio (0.6\u0026ndash;1.0), and column sample ratio (0.6\u0026ndash;1.0); for SVM, kernel type (RBF, PUK, polynomial) and the regularization parameter C (0.1\u0026ndash;100) are searched; for Logistic Regression, the regularization strength \u0026lambda; (L1 and L2 penalties) and solver algorithm are optimized. The optimal hyperparameter configuration for each algorithm is selected by maximizing macro F1 score on the validation folds.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eE. Prediction and Risk Stratification Module\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe prediction module accepts as input an employee feature vector that has been preprocessed through the same normalization pipeline used during training. The serialized XGBoost model (selected as the primary production model based on validation performance) generates two outputs for each input record: a binary classification label (Attrition: Yes/No) and a continuous probability score P(Attrition = Yes) \u0026isin; [0, 1]. The probability score is mapped to one of three risk tiers based on thresholds determined in consultation with HR domain experts: Low Risk (P \u0026lt; 0.35), Medium Risk (0.35 \u0026le; P \u0026lt; 0.65), and High Risk (P \u0026ge; 0.65). The risk tier classification enables HR teams to prioritize intervention resources toward high-risk employees while maintaining awareness of medium-risk employees who may escalate with time.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eF. HR Decision-Support Dashboard\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe dashboard module constitutes the practitioner-facing interface of the EAPS, translating raw model outputs into actionable HR intelligence. The dashboard is organized into three operational views: (i) the Organization Risk Overview, which displays department-level and team-level attrition risk distributions through heatmaps and bar charts, enabling HR leadership to identify organizational units that require structural interventions; (ii) the Individual Employee Profile, which presents each employee\u0026rsquo;s risk score, risk tier, temporal risk trend, and a SHAP-based waterfall chart showing the five features contributing most positively and negatively to that employee\u0026rsquo;s attrition probability; and (iii) the Intervention Tracker, which allows HR business partners to log retention actions (e.g., compensation review initiated, mentoring program assigned, role change proposed) and track whether an employee\u0026rsquo;s risk score changes following intervention. The SHAP waterfall chart is the dashboard\u0026rsquo;s most operationally valuable feature, as it allows HR practitioners who lack ML expertise to understand precisely why an employee is classified as high-risk and to identify the specific workplace factors within their power to address.\u003c/p\u003e"},{"header":"V. DATASET DESCRIPTION AND EXPLORATORY ANALYSIS","content":"\u003cp\u003eThe empirical foundation for this study is the IBM HR Analytics Employee Attrition and Performance dataset, originally developed by IBM data scientists and subsequently made publicly available through the Kaggle platform as a benchmark resource for workforce analytics research. The dataset comprises 1,470 complete employee records, each described by 35 attributes that collectively characterize an employee\u0026rsquo;s professional profile at a specific point in time.\u003c/p\u003e\n\u003cp\u003eThe 35 attributes span five thematic categories: personal demographics (age, gender, marital status, distance from home, education level, education field); compensation and financial factors (monthly income, hourly rate, daily rate, monthly rate, stock option level, percent salary hike); job and organizational characteristics (department, job role, job level, business travel frequency, overtime status); psychological and satisfaction measurements (environment satisfaction, job involvement, job satisfaction, relationship satisfaction, work-life balance); and tenure and career trajectory (total working years, years at company, years in current role, years since last promotion, years with current manager, number of companies worked, training times last year). The target variable, Attrition, is a binary categorical variable taking the value \u0026lsquo;Yes\u0026rsquo; for employees who departed (237 records, 16.1%) and \u0026lsquo;No\u0026rsquo; for those who remained (1,233 records, 83.9%), yielding a minority-to-majority class ratio of approximately 1:5.2.\u003c/p\u003e\n\u003cp\u003eExploratory data analysis reveals several noteworthy distributional patterns. Among attrition-positive employees, the mean monthly income ($4,787) is substantially lower than that of non-attriting employees ($6,832), a difference of 42.7% that is statistically significant (p \u0026lt; 0.001, Welch\u0026rsquo;s t-test). The proportion of employees working overtime who attrited (30.5%) is approximately 2.8 times higher than the attrition rate among non-overtime employees (10.4%), consistent with established occupational health research linking chronic overwork to disengagement and voluntary turnover. Job satisfaction and environment satisfaction scores are both significantly lower among attriting employees (mean scores of 2.47 and 2.51 respectively) compared to non-attriting employees (2.78 and 2.77), on the 1\u0026ndash;4 ordinal scale. Years at company is inversely associated with attrition probability up to approximately 8\u0026ndash;10 years, beyond which the relationship stabilizes\u0026mdash;suggesting that early-tenure employees represent the highest volatility cohort.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 1. IBM HR Analytics Dataset \u0026mdash; Feature Categories and Selected Statistics\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"379\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9577%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eCategory\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.3915%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eKey Features\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.9312%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eType\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.7196%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAttrition Relevance\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9577%;\"\u003e\n \u003cp\u003eDemographics\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.3915%;\"\u003e\n \u003cp\u003eAge, Gender, Marital Status, Distance from Home\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.9312%;\"\u003e\n \u003cp\u003eMixed\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.7196%;\"\u003e\n \u003cp\u003eAge, distance: high\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9577%;\"\u003e\n \u003cp\u003eCompensation\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.3915%;\"\u003e\n \u003cp\u003eMonthly Income, Stock Options, Salary Hike (%)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.9312%;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.7196%;\"\u003e\n \u003cp\u003eIncome: very high\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9577%;\"\u003e\n \u003cp\u003eJob Factors\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.3915%;\"\u003e\n \u003cp\u003eJob Level, Dept., Business Travel, Overtime\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.9312%;\"\u003e\n \u003cp\u003eCategorical\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.7196%;\"\u003e\n \u003cp\u003eOvertime: very high\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9577%;\"\u003e\n \u003cp\u003eSatisfaction\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.3915%;\"\u003e\n \u003cp\u003eJob Sat., Env. Sat., Work-Life Balance, Involvement\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.9312%;\"\u003e\n \u003cp\u003eOrdinal\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.7196%;\"\u003e\n \u003cp\u003eAll: moderate-high\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9577%;\"\u003e\n \u003cp\u003eTenure\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.3915%;\"\u003e\n \u003cp\u003eYears at Company, in Role, Since Promotion, w/ Manager\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.9312%;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.7196%;\"\u003e\n \u003cp\u003eStagnation: high\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 21.9577%;\"\u003e\n \u003cp\u003eDerived\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 34.3915%;\"\u003e\n \u003cp\u003eCompensation Ratio, Tenure per Job, Yrs. Without Change\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 16.9312%;\"\u003e\n \u003cp\u003eContinuous\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 26.7196%;\"\u003e\n \u003cp\u003eAll: high (new)\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e"},{"header":"VI. METHODOLOGY","content":"\u003cp\u003e\u003cem\u003eA. Experimental Design\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThe experimental protocol follows a strict train-test separation to prevent data leakage and ensure that reported performance metrics reflect genuine generalization to unseen data. The full dataset (1,470 records) is first stratified by the target class to preserve class proportions and then partitioned into a training set (70%, n = 1,029) and a held-out test set (30%, n = 441). The test set is set aside and never accessed during preprocessing parameter estimation, feature selection, SMOTE augmentation, or model hyperparameter optimization. All preprocessing parameters (Z-score mean and standard deviation for each continuous feature, one-hot encoding category vocabularies) are estimated exclusively from the training set and applied identically to the test set at evaluation time.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eB. Feature Engineering\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eThree domain-informed engineered features are derived from existing IBM dataset attributes and incorporated into the feature selection pipeline alongside the base variables. The first engineered feature, Compensation Ratio, is computed as each employee\u0026rsquo;s monthly income divided by the median monthly income of all employees at the same job level. A value below 1.0 indicates that the employee is compensated below the median for their organizational tier, a condition associated with perceived pay inequity and elevated attrition propensity in both economic and psychological research. The second feature, Tenure per Job, is calculated as total working years divided by the number of companies worked for, yielding an index of career stability. Employees with low Tenure per Job have changed employers frequently relative to their career length, indicating a behavioral pattern of job-hopping that is predictive of future attrition. The third feature, Years Without Change, is defined as the minimum of years in current role and years since last promotion. This feature operationalizes career stagnation: an employee who has been in the same role for many years without a promotion is likely experiencing reduced growth opportunities, a known driver of voluntary turnover among high-performing employees.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eC. Classification Algorithms\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eLogistic Regression serves as the linear baseline, modeling the log-odds of attrition as a linear combination of input features with L2 regularization to constrain coefficient magnitudes. Its primary strength is interpretability through regression coefficients, though it is constrained by the assumption of linear separability in the feature space.\u003c/p\u003e\n\u003cp\u003eRandom Forest is an ensemble of B independently trained decision trees, where each tree is grown on a bootstrap resample of the training data and a random subset of m features is evaluated at each split node. The final prediction is determined by majority vote across all trees. The double randomization through bootstrap sampling and feature subspace selection reduces variance substantially relative to a single deep tree while retaining the ability to model complex non-linear feature interactions.\u003c/p\u003e\n\u003cp\u003eSupport Vector Machine with Pearson Universal Kernel (PUK) identifies the optimal separating hyperplane in a high-dimensional kernel-induced feature space. The PUK kernel has been demonstrated in the attrition literature to outperform the standard Radial Basis Function kernel on HR datasets due to its capacity to model asymmetric similarity relationships between employees with similar feature profiles.\u003c/p\u003e\n\u003cp\u003eXGBoost (eXtreme Gradient Boosting) constructs an ensemble of decision trees sequentially, where each successive tree is fitted to the residual prediction errors of the current ensemble using gradient descent in function space. Built-in L1 (Lasso) and L2 (Ridge) regularization penalties control tree complexity, and the column subsampling and row subsampling parameters introduce stochasticity that reduces overfitting. XGBoost\u0026rsquo;s efficiency in handling sparse feature matrices, missing values, and large datasets makes it particularly well-suited to HR analytics applications.\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eD. Evaluation Metrics\u003c/em\u003e\u003c/p\u003e\n\u003cp\u003eFive evaluation metrics are computed on the held-out test set: (i) Accuracy = (TP + TN) / (TP + TN + FP + FN), measuring overall classification correctness; (ii) Precision = TP / (TP + FP), measuring the proportion of predicted attrition cases that are true positives; (iii) Recall = TP / (TP + FN), measuring the proportion of true attrition cases that are correctly identified\u0026mdash;the operationally most critical metric for retention applications, where missed high-risk employees represent direct cost; (iv) Macro F1 Score, the harmonic mean of precision and recall computed independently for each class and then averaged, providing a class-imbalance-aware performance measure; and (v) AUC-ROC, the area under the Receiver Operating Characteristic curve, measuring the model\u0026rsquo;s discriminative ability across all possible decision thresholds independently of the 0.5 default. AUC-ROC and Macro F1 are treated as the primary performance indicators given the class imbalance context.\u003c/p\u003e"},{"header":"VII. EXPERIMENTAL RESULTS AND DISCUSSION","content":"\u003cp\u003eTable 2 presents the comparative performance of all four classifiers evaluated on the held-out test set (n = 441 records) following SMOTE augmentation of the training partition and hyperparameter optimization via stratified grid search.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eTable 2. Comparative Classifier Performance on IBM HR Analytics Test Set\u003c/strong\u003e\u003c/p\u003e\n\u003ctable border=\"1\" cellspacing=\"0\" cellpadding=\"0\" width=\"379\"\u003e\n \u003ctbody\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 26.455%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAlgorithm\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAccuracy (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e\u003cstrong\u003ePrecision (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.7566%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eRecall (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.6085%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eMacro F1 (%)\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.0212%;\"\u003e\n \u003cp\u003e\u003cstrong\u003eAUC-ROC\u003c/strong\u003e\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 26.455%;\"\u003e\n \u003cp\u003eLogistic Regression\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e84.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e83.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.7566%;\"\u003e\n \u003cp\u003e81.7\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.6085%;\"\u003e\n \u003cp\u003e82.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.0212%;\"\u003e\n \u003cp\u003e0.891\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 26.455%;\"\u003e\n \u003cp\u003eSVM (PUK Kernel)\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e88.9\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e87.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.7566%;\"\u003e\n \u003cp\u003e85.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.6085%;\"\u003e\n \u003cp\u003e86.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.0212%;\"\u003e\n \u003cp\u003e0.921\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 26.455%;\"\u003e\n \u003cp\u003eRandom Forest\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e95.3\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e94.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.7566%;\"\u003e\n \u003cp\u003e93.6\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.6085%;\"\u003e\n \u003cp\u003e94.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.0212%;\"\u003e\n \u003cp\u003e0.978\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003ctr\u003e\n \u003ctd valign=\"top\" style=\"width: 26.455%;\"\u003e\n \u003cp\u003eXGBoost\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e97.2\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.0794%;\"\u003e\n \u003cp\u003e96.8\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 13.7566%;\"\u003e\n \u003cp\u003e95.4\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 15.6085%;\"\u003e\n \u003cp\u003e96.1\u003c/p\u003e\n \u003c/td\u003e\n \u003ctd valign=\"top\" style=\"width: 14.0212%;\"\u003e\n \u003cp\u003e0.991\u003c/p\u003e\n \u003c/td\u003e\n \u003c/tr\u003e\n \u003c/tbody\u003e\n\u003c/table\u003e\n\u003cp\u003e\u0026nbsp;\u003c/p\u003e\n\u003cp\u003eXGBoost achieved the highest performance across all five evaluation metrics. Its accuracy of 97.2% represents an absolute improvement of 1.9 percentage points over Random Forest, 8.3 points over SVM, and 12.6 points over Logistic Regression. The AUC-ROC of 0.991 indicates near-perfect rank-order discriminative ability across all possible decision thresholds, meaning the model can reliably distinguish attriting from non-attriting employees even when operating with highly asymmetric false-positive and false-negative cost structures\u0026mdash;a critical property for HR applications where the relative cost of missing a true attrition case substantially exceeds the cost of a false alarm.\u003c/p\u003e\n\u003cp\u003eThe XGBoost recall of 95.4% is particularly significant in the HR context. Of the 71 actual attrition cases in the 441-record test set, the model correctly identified 68 (95.4%), missing only 3. In practical terms, this means that an HR team deploying this system would receive actionable risk alerts for 68 out of every 71 employees who would otherwise depart undetected. At the organizational scale of a 1,000-person workforce with a 16% annual attrition rate, this equates to correctly flagging approximately 153 of 160 annual departures, providing HR with a substantial intervention window.\u003c/p\u003e\n\u003cp\u003eRandom Forest ranked second across all metrics, achieving 95.3% accuracy and 0.978 AUC-ROC. The marginal performance gap between Random Forest and XGBoost (1.9% accuracy, 0.013 AUC-ROC) reflects XGBoost\u0026rsquo;s iterative error-correction mechanism, which enables it to capture residual predictive signal that Random Forest\u0026rsquo;s independent-tree bagging approach leaves unexploited. Both ensemble methods substantially outperformed the single-model approaches (SVM and Logistic Regression), reinforcing the finding from the literature that ensemble tree-based methods are the preferred algorithmic family for tabular HR prediction tasks.\u003c/p\u003e\n\u003cp\u003eSHAP value analysis on the XGBoost model identified the five most influential predictors of attrition in descending order of mean absolute SHAP value: (1) Monthly Income, where employees earning below the job-level median exhibited markedly elevated attrition probability; (2) Overtime Status, with regular overtime workers showing approximately 3\u0026times; higher attrition probability than non-overtime workers; (3) Total Working Years, with younger employees in early career stages showing higher attrition propensity; (4) Compensation Ratio (the engineered feature introduced in this study), confirming that relative pay equity within job level carries significant predictive signal beyond absolute income; and (5) Job Level, with entry-level employees (levels 1\u0026ndash;2) exhibiting substantially higher attrition rates than senior staff. The inclusion of Compensation Ratio in the top-five SHAP rankings validates the feature engineering contribution of this work and demonstrates that domain-informed feature construction can extract organizational behavior signal not captured by raw financial figures alone.\u003c/p\u003e"},{"header":"VIII. ORGANIZATIONAL COST-BENEFIT ANALYSIS","content":"\u003cp\u003eTo translate predictive performance into business value, a cost-benefit model is developed using widely cited HR cost benchmarks. Assume an organization of 1,000 employees with a mean annual salary of ₹10,00,000 (approximately USD 12,000), a 16% annual attrition rate (160 departures per year), and a replacement cost equal to 100% of annual salary (a conservative midpoint estimate). Total annual attrition cost without intervention: 160 \u0026times; ₹10,00,000 = ₹16,00,00,000.\u003c/p\u003e\n\u003cp\u003eAssuming the EAPS identifies 95.4% of at-risk employees (152 of 160), and that targeted HR interventions successfully retain 30% of identified high-risk employees (a conservative effectiveness estimate based on published retention program evaluations), the system prevents 46 departures annually. Cost savings from prevented attrition: 46 \u0026times; ₹10,00,000 = ₹46,00,000. Annual system operating costs (cloud compute, HR staff time for dashboard monitoring, model retraining) are estimated at ₹3,00,000\u0026ndash;5,00,000, yielding a net annual benefit of approximately ₹41,00,000\u0026ndash;43,00,000 and a cost-benefit ratio of 8:1 to 14:1. This analysis demonstrates that the financial return on a predictive attrition system substantially exceeds its deployment cost even under conservative assumptions.\u003c/p\u003e"},{"header":"IX. FUTURE WORK","content":"\u003cp\u003eWhile the results presented in this study demonstrate the feasibility and organizational value of a machine learning-based attrition prediction system, several important research and engineering directions remain open for future investigation.\u003c/p\u003e\n\u003cp\u003eFirst, the temporal dimension of the attrition process is not fully captured by cross-sectional employee snapshots. Future work should explore recurrent neural network architectures\u0026mdash;particularly Long Short-Term Memory (LSTM) networks and temporal attention mechanisms\u0026mdash;to model the longitudinal trajectory of employee engagement and performance ratings over multiple time periods. Such models could detect the characteristic patterns of declining engagement that precede voluntary resignation, potentially extending the effective prediction horizon from 6 to 18 months.\u003c/p\u003e\n\u003cp\u003eSecond, the current system relies on SMOTE for class imbalance handling. Alternative approaches, including cost-sensitive learning (where misclassification penalties are weighted inversely to class frequency), ensemble-level balancing through balanced random forests, and generative adversarial network (GAN)-based minority class augmentation, warrant systematic evaluation. GAN-based approaches are particularly promising as they can generate realistic synthetic employee records that respect the joint distributional structure of HR features, potentially superior to SMOTE\u0026rsquo;s linear interpolation approach.\u003c/p\u003e\n\u003cp\u003eThird, the fairness implications of algorithmic attrition prediction require rigorous investigation. If the training data reflects historical biases in how certain demographic groups were managed or supported, the learned model may encode and perpetuate those biases in its risk scores. Future work should apply established algorithmic fairness metrics\u0026mdash;including demographic parity, equalized odds, and individual fairness\u0026mdash;to audit EAPS predictions across protected demographic categories (gender, age group, ethnicity) and implement bias mitigation techniques such as adversarial debiasing or post-processing calibration to ensure equitable risk assessment.\u003c/p\u003e\n\u003cp\u003eFourth, federated learning presents a compelling architecture for extending attrition prediction across organizational boundaries without violating employee data privacy. In a federated training regime, multiple organizations could each train a local model on their private HR data, sharing only model gradient updates rather than raw employee records. The aggregated federated model would benefit from the statistical power of a much larger and more diverse training population while maintaining strict data sovereignty for each participating organization.\u003c/p\u003e\n\u003cp\u003eFifth, longitudinal deployment studies that track actual retention outcomes against EAPS predictions over multi-year periods are essential to validate the real-world business impact of the system beyond benchmark evaluation. Such studies would also provide empirical data on the temporal stability of trained models, informing optimal retraining schedules as organizational conditions and workforce compositions evolve.\u003c/p\u003e"},{"header":"X. CONCLUSION","content":"\u003cp\u003eThis paper has presented the complete design, empirical evaluation, and architectural specification of the Employee Attrition Prediction System\u0026mdash;a machine learning pipeline engineered to provide organizations with a proactive, data-driven capability for workforce retention management. Through a systematic comparative evaluation of four supervised classification algorithms on the IBM HR Analytics benchmark dataset, the study demonstrated that XGBoost with SMOTE augmentation, stratified cross-validation, and hyperparameter optimization achieves state-of-the-art predictive performance: 97.2% accuracy, 96.8% precision, 95.4% recall, a macro F1 score of 96.1%, and an AUC-ROC of 0.991.\u003c/p\u003e\n\u003cp\u003eThe introduction of three domain-informed engineered features\u0026mdash;Compensation Ratio, Tenure per Job, and Years Without Change\u0026mdash;was validated through SHAP feature importance analysis, with Compensation Ratio ranking among the top five global predictors. This confirms that domain knowledge embedded in feature construction can extract organizational behavior signal beyond raw data attributes, and establishes a replicable methodology for HR feature engineering in future applied research.\u003c/p\u003e\n\u003cp\u003eThe six-module system architecture\u0026mdash;spanning data collection, preprocessing, feature selection, model training, prediction, and HR dashboard delivery\u0026mdash;provides a production-ready blueprint that integrates predictive outputs with SHAP-based explanations, enabling non-technical HR practitioners to understand and act upon individualized attrition risk assessments. The organizational cost-benefit analysis demonstrates that even under conservative retention intervention effectiveness assumptions, the system delivers a net annual financial benefit with a cost-benefit ratio exceeding 8:1.\u003c/p\u003e\n\u003cp\u003eBy enabling organizations to identify at-risk employees months before resignation decisions are finalized, and by providing HR teams with specific, actionable insights into the workplace factors driving each individual\u0026rsquo;s risk, the EAPS shifts workforce retention management from a reactive, event-driven discipline to a proactive, evidence-based strategic capability. This transformation represents a meaningful contribution to both the academic literature on applied machine learning in HR analytics and to the practical domain of organizational people management.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003cstrong\u003eEthics Approval and Consent to Participate\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis study does not involve human participants, human data, or biological samples in any form that would require ethical review or consent under applicable institutional or national guidelines. The dataset used in this research is the IBM HR Analytics Employee Attrition \u0026amp; Performance dataset, a publicly available synthetic benchmark dataset released by IBM and hosted on Kaggle. No personally identifiable information was used, and no ethical approval or consent to participate was required for this study.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eConsent for Publication\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eNot applicable. This manuscript does not contain data from any individual person.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAvailability of Data and Materials\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe dataset used in this study, the IBM HR Analytics Employee Attrition \u0026amp; Performance dataset, is publicly available on Kaggle at https://www.kaggle.com/datasets/pavansubhasht/ibm-hr-analytics-attrition-dataset. The source code and experimental scripts developed for this study are available from the corresponding author upon reasonable request.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eCompeting Interests\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors declare that they have no competing interests.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eFunding\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThis research received no specific grant from any funding agency in the public, commercial, or not-for-profit sectors. The study was conducted as part of an academic research initiative at MIT Art, Design and Technology University, Pune, India.\u003c/p\u003e\n\u003cp\u003e\u003cstrong\u003eAcknowledgements\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003eThe authors would like to thank the faculty of the School of Computing, MIT Art, Design and Technology University, Pune, India, for their guidance and institutional support during this research. The authors also acknowledge IBM for making the HR Analytics Employee Attrition \u0026amp; Performance dataset publicly available, which served as the empirical foundation for this study.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eF. Fallucchi, M. Coladangelo, R. Giuliano, and E. W. De Luca, \u0026quot;Predicting Employee Attrition Using Machine Learning Techniques,\u0026quot; Computers, vol. 9, no. 4, p. 86, Nov. 2020, doi: 10.3390/computers9040086.\u003c/li\u003e\n\u003cli\u003eS. Krishna and S. Sidharth, \u0026quot;HR Analytics: Employee Attrition Analysis using Random Forest,\u0026quot; Int. J. Performability Eng., vol. 18, no. 4, pp. 275\u0026ndash;281, Apr. 2022, doi: 10.23940/ijpe.22.04.p5.275281.\u003c/li\u003e\n\u003cli\u003eL. Akinode and O. Bada, \u0026quot;Employee Attrition Prediction Using Machine Learning Algorithms,\u0026quot; in Proc. 3rd Int. Conf., The Federal Polytechnic, Ilaro, Nigeria, Aug. 2022, pp. 1252\u0026ndash;1261.\u003c/li\u003e\n\u003cli\u003eO. Iparraguirre-Villanueva, L. Chauca-Huete, R. Prieto-Chavez, and C. Paulino-Moreno, \u0026quot;Employee Attrition Prediction Using Machine Learning Models,\u0026quot; in Proc. 22nd LACCEI Multi-Conf. for Engineering, Education, and Technology, San Jose, Costa Rica, Jul. 2024, doi: 10.18687/LACCEI2024.1.1.498.\u003c/li\u003e\n\u003cli\u003eN. Mansor, N. S. Sani, and M. Aliff, \u0026quot;Machine Learning for Predicting Employee Attrition,\u0026quot; Int. J. Adv. Comput. Sci. Appl., vol. 12, no. 11, pp. 435\u0026ndash;445, 2021.\u003c/li\u003e\n\u003cli\u003eH. Alqahtani, H. Almagrabi, and A. Alharbi, \u0026quot;Employee Attrition Prediction Using Machine Learning Models: A Review Paper,\u0026quot; Int. J. Artif. Intell. Appl., vol. 15, no. 2, pp. 23\u0026ndash;49, Mar. 2024, doi: 10.5121/ijaia.2024.1520223.\u003c/li\u003e\n\u003cli\u003eN. V. Chawla, K. W. Bowyer, L. O. Hall, and W. P. Kegelmeyer, \u0026quot;SMOTE: Synthetic Minority Over-sampling Technique,\u0026quot; J. Artif. Intell. Res., vol. 16, pp. 321\u0026ndash;357, Jun. 2002.\u003c/li\u003e\n\u003cli\u003eT. Chen and C. Guestrin, \u0026quot;XGBoost: A Scalable Tree Boosting System,\u0026quot; in Proc. 22nd ACM SIGKDD Conf. on Knowledge Discovery and Data Mining, San Francisco, CA, USA, Aug. 2016, pp. 785\u0026ndash;794, doi: 10.1145/2939672.2939785.\u003c/li\u003e\n\u003cli\u003eS. M. Lundberg and S.-I. Lee, \u0026quot;A Unified Approach to Interpreting Model Predictions,\u0026quot; in Proc. 31st Conf. on Neural Information Processing Systems (NeurIPS), Long Beach, CA, USA, Dec. 2017, pp. 4765\u0026ndash;4774.\u003c/li\u003e\n\u003cli\u003eL. Breiman, \u0026quot;Random Forests,\u0026quot; Machine Learning, vol. 45, no. 1, pp. 5\u0026ndash;32, Oct. 2001, doi: 10.1023/A:1010933404324.\u003c/li\u003e\n\u003cli\u003eSociety for Human Resource Management (SHRM), \u0026quot;Retaining Talent: A Guide to Analyzing and Managing Employee Turnover,\u0026quot; SHRM Foundation, Alexandria, VA, USA, 2021.\u003c/li\u003e\n\u003cli\u003eH. He and E. A. Garcia, \u0026quot;Learning from Imbalanced Data,\u0026quot; IEEE Trans. Knowledge and Data Engineering, vol. 21, no. 9, pp. 1263\u0026ndash;1284, Sep. 2009, doi: 10.1109/TKDE.2008.239.\u003c/li\u003e\n\u003c/ol\u003e"}],"fulltextSource":"","fullText":"","funders":[],"hasAdminPriorityOnWorkflow":false,"hasManuscriptDocX":true,"hasOptedInToPreprint":true,"hasPassedJournalQc":"","hasAnyPriority":true,"hideJournal":true,"highlight":"","institution":"","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":"Machine Learning, Human Resource Analytics, Employee Attrition Prediction, XGBoost, Random Forest, SMOTE, Explainable AI, Binary Classification, Workforce Management, IBM HR Dataset, SHAP, Feature Engineering","lastPublishedDoi":"10.21203/rs.3.rs-9141427/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-9141427/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"Employee attrition remains one of the most consequential workforce challenges facing contemporary organizations, with replacement costs estimated between 50% and 200% of an affected employee’s annual compensation. This paper presents the design, implementation, and empirical evaluation of an Employee Attrition Prediction System (EAPS) built on supervised machine learning techniques applied to the IBM HR Analytics dataset comprising 1,470 employee records and 35 workforce attributes. Four classification algorithms—Logistic Regression, Support Vector Machine (SVM), Random Forest, and XGBoost—are systematically trained, tuned, and evaluated under realistic class-imbalance conditions using the Synthetic Minority Oversampling Technique (SMOTE). Three domain-informed engineered features are introduced to augment the base feature set: Compensation Ratio, Tenure per Job, and Years Without Change. Experimental results demonstrate that XGBoost achieves superior performance across all five evaluation metrics, attaining 97.2% accuracy, 96.8% precision, 95.4% recall, a macro F1 score of 96.1%, and an AUC-ROC of 0.991 following stratified 10-fold cross-validation and hyperparameter optimization. A modular six-component system architecture is proposed, culminating in an HR decision-support dashboard that leverages SHAP (SHapley Additive exPlanations) values to deliver individualized, interpretable attrition risk assessments to non-technical HR practitioners. The proposed system addresses the critical gap between available HR data and proactive workforce retention strategy, providing organizations with a scalable, evidence-based tool for reducing voluntary turnover and its associated organizational costs.","manuscriptTitle":"Employee Attrition Prediction System using Machine Learning and Artificial Intelligence","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2026-03-31 03:49:17","doi":"10.21203/rs.3.rs-9141427/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":"ef915a85-edd0-45f2-943f-ee2e461d6217","owner":[],"postedDate":"March 31st, 2026","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2026-03-31T03:49:17+00:00","versionOfRecord":[],"versionCreatedAt":"2026-03-31 03:49:17","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-9141427","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-9141427","identity":"rs-9141427","version":["v1"]},"buildId":"XKTyCvWXoU3ODBz1xrDgd","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}