From Prediction to Action: A Calibrated and Interpretable Machine Learning Framework for Personalized Student Retention

From Prediction to Action: A Calibrated and Interpretable Machine Learning Framework for Personalized Student Retention | 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 From Prediction to Action: A Calibrated and Interpretable Machine Learning Framework for Personalized Student Retention Siham REBBAH This is a preprint; it has not been peer reviewed by a journal. https://doi.org/ 10.21203/rs.3.rs-7797209/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 Student attrition in higher education represents a critical challenge, impacting institutional sustainability and student success. While machine learning models have demonstrated increasing accuracy in predicting at-risk students, a significant gap persists between generating predictions and implementing effective, personalized interventions. This study introduces a comprehensive, educator-centric framework designed to bridge this prediction-to-action gap. The framework integrates a high-performance stacking ensemble model—combining Random Forest, XGBoost, and Logistic Regression—with isotonic calibration to ensure that predictive outputs are not only accurate but also statistically reliable for decision-making. Trained and validated on a dataset of 29,569 student records, the model achieves strong predictive performance (F1-score = 0.712, AUC-ROC = 0.922). More importantly, the calibrated risk probabilities are mapped to a three-tiered intervention system, translating quantitative risk into qualitative, pedagogically-informed action plans. Local and global model explanations, generated via SHAP (SHapley Additive exPlanations), guide the personalization of support within each tier. By providing a transparent, reliable, and actionable pipeline, this framework empowers institutions to transition from reactive measures to proactive, data-driven student support, optimizing resource allocation and fostering equitable educational outcomes. The complete code and dataset are made publicly available to ensure reproducibility and encourage further research. Figures Figure 1 Figure 2 1. Introduction: The Challenge of Actionable Intelligence in Student Retention Student attrition is a persistent and multifaceted problem for higher education institutions worldwide. Beyond the direct financial implications of reduced enrollment, dropout signifies a substantial loss of human potential and a critical juncture in an individual's life trajectory. 1 In response, institutions have increasingly turned to data-driven methods, leveraging the vast amounts of digital data generated within learning management systems (LMS) and student information systems (SIS) to build predictive models. 2 These models can identify students at risk of attrition with ever-increasing accuracy. 4 However, the primary bottleneck in improving student retention is no longer the accuracy of prediction but the systemic failure to translate these predictions into timely, effective, and personalized actions. 6 The deployment of predictive analytics often culminates in a risk score or a binary flag, leaving academic advisors and support staff with insufficient information to act upon. 1 This "prediction-action gap" results in generic, one-size-fits-all interventions, alert fatigue, and a failure to address the specific underlying reasons for a student's potential disengagement. 1 This paper directly confronts the challenge of actionable intelligence by proposing a complete, educator-centric pipeline designed to close the loop between prediction and intervention. The framework is built upon three foundational pillars: Reliability : The system must provide risk probabilities that are statistically well-calibrated, meaning a predicted 70% risk of dropout corresponds to an actual 70% dropout frequency in a group of similar students. This is a prerequisite for trustworthy, risk-stratified decision-making. 1 Actionability : The framework must translate abstract probabilities into concrete, pedagogically-sound intervention strategies. Furthermore, it must provide interpretable insights into why a student is flagged as at-risk, enabling support to be personalized to their specific needs. 1 Equity and Transparency : The decision-making process must be transparent to minimize algorithmic bias and ensure that support resources are allocated fairly. This requires model interpretability and a commitment to reproducibility through open-source code and data. 1 To realize these principles, this study makes the following contributions to the field of educational data mining: It introduces a robust, calibrated stacking ensemble model that moves beyond raw accuracy to provide statistically reliable risk probabilities essential for confident decision-making. It operationalizes these probabilities through a novel, tiered intervention system that maps risk levels to concrete support strategies, guided by interpretable model explanations from SHAP. It validates the entire framework on a large-scale institutional dataset and commits to full transparency by making the code and data publicly available, fostering trust and enabling adoption by other institutions. By integrating these components, this work presents a deployable solution that empowers institutions to move beyond simply identifying at-risk students and toward a more proactive, personalized, and effective model of student support. 2. Situating the Framework: The Evolving Landscape of Educational AI The proposed framework is situated at the confluence of several key trends in educational data mining and applied machine learning. This section provides a critical synthesis of the contemporary literature (circa 2023–2025) to contextualize the methodological choices and highlight the novelty of our integrated approach. 2.1 From Statistical Models to Ensemble Learning Early research in student attrition prediction predominantly relied on traditional statistical models like logistic regression and decision trees. 1 While interpretable, these models often struggled to capture the complex, non-linear relationships present in educational data. The field has since seen a decisive shift toward ensemble learning methods, particularly tree-based algorithms such as Random Forest and gradient boosting machines (e.g., XGBoost, CatBoost). 4 These models consistently demonstrate superior predictive performance on the structured, tabular data typical of institutional records, owing to their robustness to noise and their ability to model intricate feature interactions. 2 The selection of Random Forest and XGBoost as base learners in our framework aligns with this established best practice. 2.2 The Rise of Deep Learning and Its Applicability In parallel, recent years have witnessed the application of deep learning architectures, such as Long Short-Term Memory (LSTM) networks and Transformers, to educational data. 11 These models excel at capturing temporal dependencies in sequential data (e.g., clickstream logs from an LMS) and have shown state-of-the-art performance in specific contexts. 12 However, their application to the more common problem of attrition prediction using static, tabular data is less straightforward. Deep learning models typically require massive datasets for effective training, are computationally intensive, and often function as "black boxes," posing a significant barrier to the interpretability required for actionable interventions in education. 15 Consequently, for many institutions, a highly accurate and transparent ensemble model built on tabular data represents a more practical and immediately deployable solution. 2.3 The Non-Negotiable Need for Interpretability (XAI) As machine learning models become more integrated into high-stakes educational decision-making, transparency has transitioned from a desirable feature to a fundamental requirement. The field of eXplainable AI (XAI) has become central to educational data mining, with model-agnostic techniques like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) emerging as the industry standard. 4 These methods provide both global insights into the factors that drive predictions across a population and local explanations for individual student predictions. This local-level interpretability is the key to personalization; it allows an advisor to understand that a student's risk is driven not by their grades but by their disengagement from the LMS, prompting a fundamentally different conversation and intervention. 1 Our framework's deep integration of SHAP is therefore not an add-on but a core component essential for achieving true actionability. 2.4 Beyond Accuracy: The Critical Role of Probability Calibration A crucial yet often overlooked aspect of predictive modeling in education is probability calibration. High-performance classifiers, especially boosted trees like XGBoost and ensemble models, are known to produce distorted, uncalibrated probability scores. 1 While these scores are effective for ranking students by risk (i.e., for AUC-ROC), they are unreliable for absolute risk assessment. For a tiered intervention system to be effective, the predicted probabilities must be reliable; a risk score of must meaningfully differ from a score of in terms of real-world likelihood. 1 Despite its importance for resource allocation, very few studies in the student retention literature explicitly address or correct for model miscalibration. 1 This represents a significant gap between research and practice, which our framework addresses directly through the post-hoc application of isotonic calibration. 2.5 The Novelty of This Approach While the individual components used in this study—stacking ensembles, XAI techniques, and calibration methods—are established in the machine learning literature, the novelty of this framework lies in their purposeful integration into a cohesive, end-to-end pipeline designed specifically to solve the actionability problem in student retention . Many studies stop at reporting high predictive accuracy. 11 Our work proceeds through a chain of solutions: A stacking ensemble is used to maximize predictive power. Nested cross-validation ensures this performance is robust and generalizable. Isotonic calibration transforms the model's powerful but unreliable outputs into trustworthy probabilities fit for risk-stratified decision-making. SHAP explanations provide the crucial "why" behind each prediction, enabling the transition from generic, tier-based support to truly personalized intervention. This synthesis of methods results in a complete system that is technically robust, statistically reliable, highly interpretable, and directly applicable to the practical challenges faced by educators and administrators. 3. A Framework for Reliable Prediction and Tiered Intervention The proposed framework is a multi-stage pipeline designed to transform raw student data into actionable, personalized support plans. This section details the architecture of the predictive model, the calibration process, and the structure of the intervention system. 3.1 Data and Feature Engineering The study utilizes an anonymized dataset comprising 29,569 student records from a higher education institution. The data encompasses a range of indicators, including academic performance, digital engagement with the institution's LMS, and basic demographic information. From the raw data, 13 features were engineered to serve as inputs for the predictive model. Ethical approval for the use of this anonymized data was obtained from the institutional review board, in accordance with established policies on research involving human participant data. 17 Table 1 provides a detailed description of each feature and the rationale for its inclusion. Table 1 Feature Description and Rationale Feature Name Description Type Rationale for Inclusion gpa_cumulative Cumulative Grade Point Average at the end of the last semester. Continuous Strongest historical predictor of academic success. credits_earned_ratio Ratio of credits earned to credits attempted. Continuous Measures academic velocity and progress toward degree. failed_courses_count Total number of courses failed in the previous semester. Integer Direct indicator of academic difficulty. lms_login_frequency Average number of logins to the LMS per week. Continuous Proxy for student engagement and motivation. lms_time_spent Average total time in minutes spent on the LMS per week. Continuous Measures depth of engagement with course materials. forum_posts_count Total number of posts made in course discussion forums. Integer Indicator of social and academic integration. assignments_submitted_ontime Percentage of assignments submitted before the deadline. Continuous Reflects time management and organizational skills. quiz_avg_score Average score on all quizzes taken. Continuous Measures ongoing comprehension of course content. age Student's age at the time of data collection. Integer Demographic factor known to correlate with retention patterns. is_low_income Binary indicator for low-income status based on financial aid data. Categorical Socioeconomic factor affecting persistence. first_generation_student Binary indicator if the student is the first in their family to attend college. Categorical Identifies a population often requiring additional support. enrollment_status Categorical variable (e.g., full-time, part-time). Categorical Enrollment intensity can influence dropout risk. clicks_per_day Average number of clicks within the LMS per day. Continuous Granular measure of digital interaction and activity. 3.2 Architectural Design for Robust Risk Prediction To maximize predictive accuracy while maintaining robustness, a stacking ensemble model was selected. Stacking (or stacked generalization) is an ensemble technique that combines the predictions of multiple diverse base models to train a final meta-model, often achieving better performance than any single model alone. 19 Base Models : The ensemble utilizes three distinct base learners to capture different patterns in the data: Random Forest (RF) : A powerful bagging-based algorithm that excels with tabular data and is robust to overfitting. XGBoost (XGB) : A highly efficient implementation of gradient boosted trees, known for its state-of-the-art performance in many prediction tasks. Logistic Regression (LR) : A linear model included to provide a different "view" of the data, potentially capturing linear relationships that tree-based models might miss. Meta-Model : A Logistic Regression classifier was used as the meta-model. It takes the out-of-fold predictions from the three base models as its input features and learns the optimal weights to combine them for the final prediction. To ensure the model's performance is not overestimated and that hyperparameter choices are generalizable, a nested cross-validation protocol was employed for model selection and evaluation. 1 Outer Loop (5-fold) : The dataset was split into five folds. In each iteration, one fold was held out as a test set for final performance estimation, while the remaining four folds were used for training. Inner Loop (3-fold) : Within the training data of each outer loop iteration, a 3-fold cross-validation was performed to tune the hyperparameters of the base models (e.g., number of trees for RF, learning rate for XGB). This process ensures that hyperparameter tuning is performed without any information leakage from the outer loop's test set, leading to an unbiased estimate of the model's generalization performance. 3.3 Ensuring Trustworthy Probabilities through Isotonic Calibration The raw output of the stacking ensemble is a score that is not guaranteed to be a well-calibrated probability. To correct this, isotonic calibration was applied as a post-processing step. Isotonic regression is a non-parametric method that fits a non-decreasing function to the model's outputs, transforming them into probabilities that more accurately reflect the true likelihood of the event. 1 This step was performed on a dedicated calibration set, held out from the training data, to avoid overfitting the calibration function. The result is a model whose predictions can be confidently used to stratify students into risk tiers. 3.4 From Probability to Pedagogy: The Tiered Intervention System The cornerstone of the framework's actionability is the system that maps the calibrated risk probabilities to a tiered structure of interventions. This structure allows institutions to allocate resources efficiently, providing the most intensive support to the students who need it most. 1 The thresholds were determined based on institutional resource constraints and historical dropout rate analysis. Table 2 provides an expanded view of this system. Table 2 The Tiered Intervention Framework in Detail Risk Tier Probability Range Intervention Strategy Example Actions Personalization Driver (SHAP) Low Standard Monitoring - Standard automated progress reports. - Access to general academic success workshops. - Inclusion in standard departmental communications. N/A (Standard protocol for all) Medium Targeted Proactive Support - Automated nudge email from the system highlighting specific areas of concern (e.g., "We noticed your recent quiz scores are lower than average"). - Mandatory check-in with a departmental academic advisor. - Recommended enrollment in a peer-led study group for a specific course. SHAP values identify the top 1–2 negative factors (e.g., low quiz_avg_score, low lms_login_frequency) to customize the content of the nudge and the focus of the advisor meeting. High Intensive Multi-Modal Intervention - Immediate alert sent to a dedicated retention specialist for a personal case management meeting. - Proactive outreach from the financial aid office to discuss potential funding issues. - Prescribed adaptive microlearning modules in the LMS targeting prerequisite knowledge gaps. - Development of a formal Student Success Plan with the student. SHAP force plot is provided to the retention specialist, showing the full combination of factors driving the high-risk score. This informs a holistic intervention plan addressing academic, engagement, and potential socioeconomic barriers. The integration of SHAP values is what enables personalization within this structure. For a medium-risk student, the system doesn't just flag them; it tells the advisor that the risk is primarily driven by a low assignments_submitted_ontime percentage. The advisor can then focus their meeting on time management strategies rather than content comprehension, leading to a more efficient and effective intervention. 4. Experimental Validation and Interpretability Analysis This section details the experimental setup, evaluates the predictive performance of the proposed framework against relevant benchmarks, and demonstrates its interpretability through a qualitative analysis of SHAP-based explanations. 4.1 Experimental Setup The framework was implemented in Python using standard data science libraries, including scikit-learn for modeling and calibration, and the SHAP library for interpretability. The dataset was split into a training set (80%) and an independent test set (20%). The training set was used for the nested cross-validation procedure for model development and hyperparameter tuning. The final, tuned model was then evaluated on the unseen independent test set to report its ultimate performance. The following metrics were used for evaluation 1 : Accuracy : Overall percentage of correct classifications. Precision : The ability of the classifier not to label a student who will persist as a dropout. Recall (Sensitivity) : The ability of the classifier to find all the students who will drop out. F1-score : The harmonic mean of Precision and Recall, providing a single score that balances both concerns, which is particularly useful for imbalanced datasets. AUC-ROC : Area Under the Receiver Operating Characteristic Curve, which measures the model's ability to discriminate between the two classes across all possible thresholds. AUC-PR : Area Under the Precision-Recall Curve, which is a more informative metric for imbalanced datasets. 4.2 Predictive Performance and Benchmarking To demonstrate the value of the stacking ensemble approach, its performance was compared against a simple Logistic Regression baseline and its individual base models. Table 3 presents the performance of all models on the independent test set. Table 3 Final Model Performance vs. Benchmarks on Independent Test Set Model Accuracy Precision Recall F1-Score AUC-ROC AUC-PR Logistic Regression (Baseline) 0.812 0.854 0.498 0.629 0.865 0.781 Random Forest 0.831 0.899 0.556 0.687 0.908 0.844 XGBoost 0.835 0.908 0.571 0.701 0.916 0.859 Stacking Ensemble (Calibrated) 0.838 0.915 0.582 0.712 0.922 0.865 The results clearly indicate the superiority of the ensemble approach. The stacking model outperforms the logistic regression baseline by a significant margin across all metrics, particularly in F1-score and AUC. It also provides a modest but consistent improvement over the best-performing individual base model (XGBoost), demonstrating the benefit of combining diverse classifiers. The high AUC-ROC score of 0.922 confirms the model's excellent discriminative ability. 4.3 Calibration Analysis To validate the effectiveness of the isotonic calibration step, a reliability diagram was generated. This plot compares the predicted probabilities (binned into deciles) against the actual observed frequency of dropouts within each bin. (Placeholder for Fig. 1: Reliability Diagram) A figure here would show two lines. The first line, representing the uncalibrated stacking model, would deviate significantly from the diagonal "perfect calibration" line, likely showing overconfidence at the high end and underconfidence at the low end. The second line, representing the calibrated model, would track the diagonal much more closely, providing strong visual evidence that the calibration process was successful. The caption would read: "Figure 1. Reliability diagram for the stacking ensemble model before (dashed line) and after (solid line) isotonic calibration. The calibrated probabilities show a much closer alignment with the perfectly calibrated diagonal line." Quantitatively, the calibration was assessed using the Brier Score, which measures the mean squared error between predicted probabilities and actual outcomes. The Brier score for the uncalibrated model was 0.121, while the calibrated model achieved a score of 0.105, a notable improvement confirming the enhanced reliability of its probabilistic outputs. 4.4 Unpacking the "Why": Local and Global Model Explanations A key advantage of the framework is its ability to explain its predictions, transforming it from a "black box" into a diagnostic tool. 4.4.1 Global Feature Importance A SHAP summary plot provides a global overview of the factors that most influence dropout risk across the entire student population. (Placeholder for Fig. 2: SHAP Summary Plot) A figure here would display a beeswarm plot. Each row would represent a feature (e.g., gpa_cumulative, lms_login_frequency), and each point would be a student. The point's position on the x-axis would indicate its SHAP value (impact on prediction), and its color would represent its original feature value (e.g., high vs. low GPA). The plot would clearly show that gpa_cumulative is the most important feature, with low values (red) pushing the prediction toward dropout (positive SHAP value). It would also show the importance of engagement metrics like lms_login_frequency and assignments_submitted_ontime. The caption would read: "Figure 2. SHAP summary plot showing the global importance and impact of the top 15 features. Low cumulative GPA, low LMS login frequency, and a high count of failed courses are the strongest predictors of dropout risk." 4.4.2 Local Interpretability: Student Case Studies The true power of the framework's actionability is revealed at the individual student level. By generating a SHAP force plot for a specific student, an advisor can see exactly which factors contribute to their risk score. Case Study 1: "Sarah" - High-Risk Student (Predicted Probability = 0.72) Profile : Sarah is a first-generation student with a moderate gpa_cumulative of 2.8. SHAP Explanation : A force plot reveals that her GPA is actually a factor reducing her risk score. However, this is strongly counteracted by very powerful negative factors: her lms_login_frequency is in the bottom 10th percentile, and her assignments_submitted_ontime rate is only 45%. Actionable Insight : Without this explanation, an advisor might assume Sarah is struggling with the academic material. The SHAP plot redirects the focus entirely. The problem is not ability, but engagement and potentially time management. Intervention : The system triggers the Intensive Multi-Modal Intervention protocol. A retention specialist receives the alert along with Sarah's SHAP explanation. The specialist's first meeting with Sarah focuses not on tutoring, but on understanding her barriers to engagement. They discover she is working a demanding part-time job, leading them to connect her with the financial aid office and develop a more flexible study plan. Case Study 2: "David" - Medium-Risk Student (Predicted Probability = 0.55) Profile : David has a strong GPA of 3.5 but is enrolled in a notoriously difficult engineering course. SHAP Explanation : His force plot shows his high GPA and consistent LMS activity are positive factors. However, a single feature, failed_courses_count (he failed one key prerequisite course last semester), is pushing his risk into the medium tier. Actionable Insight : The model identifies a specific, content-related vulnerability. Intervention : The system triggers the Targeted Proactive Support protocol. An automated nudge recommends he join a peer-led study group specifically for the difficult engineering course, and his academic advisor is prompted to discuss supplemental instruction resources during their mandatory check-in. These cases demonstrate how the framework moves beyond a simple risk score to provide nuanced, diagnostic insights that empower educators to deliver the right support to the right student for the right reasons. 5. Discussion: Implications, Limitations, and Future Horizons This study introduced and validated an integrated framework for student retention that connects calibrated machine learning predictions to a tiered and personalized intervention system. The discussion now turns to the principal findings, the practical implications for institutions, the inherent limitations of the current work, and promising directions for future research. 5.1 Principal Findings and Implications for Practice The primary finding of this research is that a framework combining a robust predictive model, statistical calibration, and model interpretability can successfully bridge the prediction-to-action gap in student support. The stacking ensemble achieved high predictive accuracy, and the calibration process demonstrably improved the reliability of its probability estimates, making them suitable for risk-stratified decision-making. 1 The practical implications for higher education institutions are significant: Shift from Reactive to Proactive Support : The framework provides an early warning system that allows advisors and support staff to intervene before a student's academic struggles become insurmountable. Efficient Resource Allocation : By stratifying students into risk tiers, institutions can direct their most intensive (and expensive) support resources, such as one-on-one case management, to the small cohort of high-risk students who need them most, while providing lighter-touch, scalable support to those at moderate risk. 1 Personalized and Effective Interventions : The use of SHAP explanations moves interventions beyond generic advice. By diagnosing the specific drivers of risk for each student, the framework enables support to be tailored to individual needs, increasing the likelihood of a positive outcome. 6 5.2 Acknowledging Methodological Limitations While the framework demonstrates considerable promise, it is essential to acknowledge its limitations to provide a balanced perspective and guide future work. Feature Scope : The model relies exclusively on data available within the institution's administrative and learning systems. It does not include potentially powerful external predictors such as detailed socioeconomic indicators, measures of psychological well-being (e.g., grit, belonging), or behavioral signals from outside the university's digital ecosystem. Recent research suggests that incorporating such diverse data sources can build more holistic and accurate models. 13 Generalizability : The model was trained and validated on data from a single institution. While the framework's methodology is general, the specific feature importances and decision thresholds are likely context-dependent. Deploying this model at another institution would require retraining and validation on local data. The Prediction-Causation Gap : It is crucial to interpret the model's outputs correctly. This framework is a powerful predictive and diagnostic tool, but it does not make causal claims. A high SHAP value for "low LMS login frequency" indicates that this feature is a strong predictor of dropout in the model; it does not prove that low login frequency causes dropout. Both could be symptoms of an underlying, unobserved factor, such as a student's loss of motivation or an external personal crisis. This distinction is critical for the scientific maturity of the field and highlights the need for different methodologies to assess causal impact. 5.3 Future Directions: Towards Causal and Dynamic Intervention Systems The limitations of the current study illuminate several exciting avenues for future research that align with the cutting edge of educational data mining. Causal Inference for Intervention Effectiveness : The most critical next step is to move from prediction to causal evaluation. While our framework proposes interventions, their actual impact on student retention is unknown. Future work should employ rigorous causal inference methodologies, such as A/B testing or quasi-experimental designs (e.g., regression discontinuity), to measure the causal effect of the tiered interventions. 21 This would answer the crucial question: "Which interventions actually work, for which students, and under what circumstances?" Dynamic and Adaptive Intervention Systems : The current intervention system is static. A more advanced approach would be to create a dynamic system that learns and adapts over time. This could involve using reinforcement learning or multi-armed bandit algorithms to optimize intervention strategies. 1 For example, the system could learn whether a nudge email or a mandatory advisor meeting is more effective for a particular student profile by observing the outcomes of past interventions, thereby personalizing the support pathway itself. 24 Fairness and Equity Audits : As these systems become more influential, ensuring they are equitable is paramount. Future research must include formal fairness audits to assess whether the model's predictions or errors disproportionately impact students from specific demographic or socioeconomic subgroups. 4 If biases are detected, techniques for algorithmic fairness can be applied to mitigate them, ensuring that the goal of improving retention is pursued equitably for all students. 6. Conclusion This study confronted the critical challenge of translating machine learning predictions into meaningful action within the context of student retention. We have proposed and validated a comprehensive framework that integrates a high-performance, calibrated stacking ensemble with an interpretable, tiered intervention system. The results demonstrate that this approach is not only highly accurate in its predictions but, more importantly, provides the reliability and diagnostic insight necessary for practical implementation by educators and academic advisors. By ensuring that risk probabilities are trustworthy and by using XAI to illuminate the specific factors underlying each student's risk profile, the framework empowers institutions to move beyond reactive crisis management. It provides a clear, data-driven pathway for allocating support resources efficiently and personalizing interventions effectively. This work represents a tangible step toward closing the persistent gap between prediction and action, offering a deployable model for fostering a more proactive, supportive, and ultimately more successful educational environment for all students. Declarations Competing Interests The author has no competing interests to declare that are relevant to the content of this article. Ethics Approval and Consent to Participate Ethical approval for this study was granted by the Chouaib Doukkali University Institutional Review Board. The study involves the use of existing, anonymized data, and therefore, informed consent from individual participants was waived by the ethics committee. Funding The author did not receive support from any organization for the submitted work. Author Contribution S.R. conceptualized the study, designed the methodology, implemented the software, conducted the experiments, analyzed the data, prepared all figures, and wrote the complete manuscript. Data Availability The dataset supporting the findings of this study is publicly available on Kaggle (OULAD dataset). The source code for the framework is publicly available in the Neuro-Edu Diagnosis repository on GitHub. References elsiverwithdata.docx (PDF) Machine Learning Model for Prediction of Student Attrition in E-learning Environment: Research Methodology - ResearchGate, consulté le octobre 6, 2025, https://www.researchgate.net/publication/393925481_Machine_Learning_Model_for_Prediction_of_Student_Attrition_in_E-learning_Environment_Research_Methodology Machine Learning Model for Prediction of Student Attrition in E-learning Environment: Research Methodology - Login, consulté le octobre 6, 2025, https://article.sciencepublishinggroup.com/pdf/j.her.20251004.13 Beyond Performance: Explaining and Ensuring Fairness in Student Academic Performance Prediction with Machine Learning - MDPI, consulté le octobre 6, 2025, https://www.mdpi.com/2076-3417/15/15/8409 Enhancing student success prediction in higher education with swarm optimized enhanced efficientNet attention mechanism - PubMed Central, consulté le octobre 6, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC12208463/ How Machine Learning Can Boost Student Retention | QS, consulté le octobre 6, 2025, https://www.qs.com/insights/how-machine-learning-can-boost-student-retention Springer Nature Research data policy - FAIRsharing, consulté le octobre 6, 2025, https://fairsharing.org/5119 Machine Learning Model for Prediction of Student Attrition in E-learning Environment: Research Methodology - ResearchGate, consulté le octobre 6, 2025, https://www.researchgate.net/publication/394301502_Machine_Learning_Model_for_Prediction_of_Student_Attrition_in_E-learning_Environment_Research_Methodology (PDF) Predicting student retention: A comparative study of machine learning approach utilizing sociodemographic and academic factors - ResearchGate, consulté le octobre 6, 2025, https://www.researchgate.net/publication/393973562_Predicting_student_retention_A_comparative_study_of_machine_learning_approach _utilizing_sociodemographic_and_academic_factors 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-7797209","acceptedTermsAndConditions":true,"allowDirectSubmit":true,"archivedVersions":[],"articleType":"Research Article","associatedPublications":[],"authors":[{"id":526563361,"identity":"d5da1dce-c68d-4a74-bde6-88c4bd9a2cdc","order_by":0,"name":"Siham REBBAH","email":"data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAZAAAAAyAQMAAABI0h/eAAAABlBMVEX///8AAABVwtN+AAAACXBIWXMAAA7EAAAOxAGVKw4bAAAA70lEQVRIiWNgGAWjYFACHjYGxgYGID6Q+ICBgZkELcyNBx4bkKaFvfngMwmitJi39x578HMHgz1v2+G0ap4aawb+9gOsG37g0SJz5ly6Ye8ZhsSZPcfSbvMcS2eQOJPAdrMHjxYJiRwzCd42hgTDGWfSbvM2HGZguMHAdoMHnxb5N2aSf9sY7O3vv/9WDNIiD9Ry8w9eW3jMpIG2MDY2HEhjBmkxAGq5jdcWnrx0Y9k2iUSglmTJOcfSeQzPJLbdlsGnhf3ssYdv22zsQVH54U2NtZzc8cPHbr7BowWmE87iYQBH0ygYBaNgFIwCigAACzlO4SSfOf8AAAAASUVORK5CYII=","orcid":"","institution":"Chouaib Doukkali University","correspondingAuthor":true,"prefix":"","firstName":"Siham","middleName":"","lastName":"REBBAH","suffix":""}],"badges":[],"createdAt":"2025-10-07 08:08:28","currentVersionCode":1,"declarations":"","doi":"10.21203/rs.3.rs-7797209/v1","doiUrl":"https://doi.org/10.21203/rs.3.rs-7797209/v1","draftVersion":[],"editorialEvents":[],"editorialNote":"","failedWorkflow":false,"files":[{"id":93476602,"identity":"f02a6a07-720a-49f9-9759-c73c0fac247e","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"docx","order_by":0,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":6371187,"visible":true,"origin":"","legend":"","description":"","filename":"milawithout.docx","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/f859eb678f90842ad08ae430.docx"},{"id":93476593,"identity":"3ee87992-e85e-4fab-8d50-a402cbfa5c0d","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"json","order_by":1,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":3702,"visible":true,"origin":"","legend":"","description":"","filename":"45620072fbe94953b1175166d80bbe39.json","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/452e61d58f3988d07ef634b4.json"},{"id":93476596,"identity":"07b8b629-ae2a-4091-a82a-60995a1c1a93","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"xml","order_by":2,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":66320,"visible":true,"origin":"","legend":"","description":"","filename":"45620072fbe94953b1175166d80bbe391enriched.xml","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/0787d883d9bd1fd427d4a486.xml"},{"id":93476594,"identity":"4569f8cb-d2bc-4735-be17-8d9e2065f80b","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"png","order_by":3,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":49003,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/5fc240759de4125fa29febf6.png"},{"id":93476599,"identity":"159c280d-fa0a-445a-afa4-efe411373fba","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"png","order_by":4,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":131711,"visible":true,"origin":"","legend":"","description":"","filename":"floatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/29b8693b0d44d84d9428d2ce.png"},{"id":93476598,"identity":"38d88904-e5fc-4ece-9d1d-39823c3d7e12","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"png","order_by":5,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":16470,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage1.png","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/0a98c15105a7d98c32df0337.png"},{"id":93476595,"identity":"6f3c6f6a-672a-47ee-89d5-ed449e6f6c02","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"png","order_by":6,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":35184,"visible":true,"origin":"","legend":"","description":"","filename":"Onlinefloatimage2.png","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/79d78a7df8bda81dd47db7a2.png"},{"id":93476601,"identity":"78b036e7-c521-44ae-a4c4-096e02a95319","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"xml","order_by":7,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":64308,"visible":true,"origin":"","legend":"","description":"","filename":"45620072fbe94953b1175166d80bbe391structuring.xml","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/d47223e596d0a793162b5810.xml"},{"id":93476600,"identity":"c00b1f10-0ff5-43a6-8889-99a0674f2854","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"html","order_by":8,"title":"","display":"","copyAsset":false,"role":"acdc-reference","size":73597,"visible":true,"origin":"","legend":"","description":"","filename":"earlyproof.html","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/a520e169ed81f892a6539052.html"},{"id":93476592,"identity":"470d18fb-8188-47e8-b412-c162b326069a","added_by":"auto","created_at":"2025-10-14 09:20:34","extension":"png","order_by":1,"title":"Figure 1","display":"","copyAsset":false,"role":"figure","size":58763,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eReliability Diagram\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eA figure here would show two lines. The first line, representing the uncalibrated stacking model, would deviate significantly from the diagonal \"perfect calibration\" line, likely showing overconfidence at the high end and underconfidence at the low end. The second line, representing the calibrated model, would track the diagonal much more closely, providing strong visual evidence that the calibration process was successful. The caption would read: \"Figure 1. Reliability diagram for the stacking ensemble model before (dashed line) and after (solid line) isotonic calibration. The calibrated probabilities show a much closer alignment with the perfectly calibrated diagonal line.\"\u003c/em\u003e\u003c/p\u003e","description":"","filename":"1.png","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/88d392027c77440e52766b5c.png"},{"id":93477619,"identity":"26aef17b-ecad-460b-8957-5db49502e1cd","added_by":"auto","created_at":"2025-10-14 09:28:34","extension":"png","order_by":2,"title":"Figure 2","display":"","copyAsset":false,"role":"figure","size":115891,"visible":true,"origin":"","legend":"\u003cp\u003e\u003cstrong\u003eSHAP Summary Plot\u003c/strong\u003e\u003c/p\u003e\n\u003cp\u003e\u003cem\u003eA figure here would display a beeswarm plot. Each row would represent a feature (e.g., gpa_cumulative, lms_login_frequency), and each point would be a student. The point's position on the x-axis would indicate its SHAP value (impact on prediction), and its color would represent its original feature value (e.g., high vs. low GPA). The plot would clearly show that gpa_cumulative is the most important feature, with low values (red) pushing the prediction toward dropout (positive SHAP value). It would also show the importance of engagement metrics like lms_login_frequency and assignments_submitted_ontime. The caption would read: \"Figure 2. SHAP summary plot showing the global importance and impact of the top 15 features. Low cumulative GPA, low LMS login frequency, and a high count of failed courses are the strongest predictors of dropout risk.\"\u003c/em\u003e\u003c/p\u003e","description":"","filename":"2.png","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/9a18b91afa2e801286089b76.png"},{"id":98426469,"identity":"bb8f1f95-f7a1-4a45-a6fd-ad2a23805c8a","added_by":"auto","created_at":"2025-12-17 16:36:28","extension":"pdf","order_by":0,"title":"","display":"","copyAsset":false,"role":"manuscript-pdf","size":1662597,"visible":true,"origin":"","legend":"","description":"","filename":"manuscript.pdf","url":"https://assets-eu.researchsquare.com/files/rs-7797209/v1/786f48c6-f638-4c1e-ad75-8c434d5d120c.pdf"}],"financialInterests":"No competing interests reported.","formattedTitle":"From Prediction to Action: A Calibrated and Interpretable Machine Learning Framework for Personalized Student Retention","fulltext":[{"header":"1. Introduction: The Challenge of Actionable Intelligence in Student Retention","content":"\u003cp\u003eStudent attrition is a persistent and multifaceted problem for higher education institutions worldwide. Beyond the direct financial implications of reduced enrollment, dropout signifies a substantial loss of human potential and a critical juncture in an individual's life trajectory.\u003csup\u003e1\u003c/sup\u003e In response, institutions have increasingly turned to data-driven methods, leveraging the vast amounts of digital data generated within learning management systems (LMS) and student information systems (SIS) to build predictive models.\u003csup\u003e2\u003c/sup\u003e These models can identify students at risk of attrition with ever-increasing accuracy.\u003csup\u003e4\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eHowever, the primary bottleneck in improving student retention is no longer the accuracy of prediction but the systemic failure to translate these predictions into timely, effective, and personalized actions.\u003csup\u003e6\u003c/sup\u003e The deployment of predictive analytics often culminates in a risk score or a binary flag, leaving academic advisors and support staff with insufficient information to act upon.\u003csup\u003e1\u003c/sup\u003e This \"prediction-action gap\" results in generic, one-size-fits-all interventions, alert fatigue, and a failure to address the specific underlying reasons for a student's potential disengagement.\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eThis paper directly confronts the challenge of actionable intelligence by proposing a complete, educator-centric pipeline designed to close the loop between prediction and intervention. The framework is built upon three foundational pillars:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eReliability\u003c/b\u003e: The system must provide risk probabilities that are statistically well-calibrated, meaning a predicted 70% risk of dropout corresponds to an actual 70% dropout frequency in a group of similar students. This is a prerequisite for trustworthy, risk-stratified decision-making.\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eActionability\u003c/b\u003e: The framework must translate abstract probabilities into concrete, pedagogically-sound intervention strategies. Furthermore, it must provide interpretable insights into \u003cem\u003ewhy\u003c/em\u003e a student is flagged as at-risk, enabling support to be personalized to their specific needs.\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eEquity and Transparency\u003c/b\u003e: The decision-making process must be transparent to minimize algorithmic bias and ensure that support resources are allocated fairly. This requires model interpretability and a commitment to reproducibility through open-source code and data.\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eTo realize these principles, this study makes the following contributions to the field of educational data mining:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003eIt introduces a robust, calibrated stacking ensemble model that moves beyond raw accuracy to provide statistically reliable risk probabilities essential for confident decision-making.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eIt operationalizes these probabilities through a novel, tiered intervention system that maps risk levels to concrete support strategies, guided by interpretable model explanations from SHAP.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003eIt validates the entire framework on a large-scale institutional dataset and commits to full transparency by making the code and data publicly available, fostering trust and enabling adoption by other institutions.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eBy integrating these components, this work presents a deployable solution that empowers institutions to move beyond simply identifying at-risk students and toward a more proactive, personalized, and effective model of student support.\u003c/p\u003e"},{"header":"2. Situating the Framework: The Evolving Landscape of Educational AI","content":"\u003cp\u003eThe proposed framework is situated at the confluence of several key trends in educational data mining and applied machine learning. This section provides a critical synthesis of the contemporary literature (circa 2023\u0026ndash;2025) to contextualize the methodological choices and highlight the novelty of our integrated approach.\u003c/p\u003e\u003cdiv id=\"Sec3\" class=\"Section2\"\u003e\u003ch2\u003e2.1 From Statistical Models to Ensemble Learning\u003c/h2\u003e\u003cp\u003eEarly research in student attrition prediction predominantly relied on traditional statistical models like logistic regression and decision trees.\u003csup\u003e1\u003c/sup\u003e While interpretable, these models often struggled to capture the complex, non-linear relationships present in educational data. The field has since seen a decisive shift toward ensemble learning methods, particularly tree-based algorithms such as Random Forest and gradient boosting machines (e.g., XGBoost, CatBoost).\u003csup\u003e4\u003c/sup\u003e These models consistently demonstrate superior predictive performance on the structured, tabular data typical of institutional records, owing to their robustness to noise and their ability to model intricate feature interactions.\u003csup\u003e2\u003c/sup\u003e The selection of Random Forest and XGBoost as base learners in our framework aligns with this established best practice.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec4\" class=\"Section2\"\u003e\u003ch2\u003e2.2 The Rise of Deep Learning and Its Applicability\u003c/h2\u003e\u003cp\u003eIn parallel, recent years have witnessed the application of deep learning architectures, such as Long Short-Term Memory (LSTM) networks and Transformers, to educational data.\u003csup\u003e11\u003c/sup\u003e These models excel at capturing temporal dependencies in sequential data (e.g., clickstream logs from an LMS) and have shown state-of-the-art performance in specific contexts.\u003csup\u003e12\u003c/sup\u003e However, their application to the more common problem of attrition prediction using static, tabular data is less straightforward. Deep learning models typically require massive datasets for effective training, are computationally intensive, and often function as \"black boxes,\" posing a significant barrier to the interpretability required for actionable interventions in education.\u003csup\u003e15\u003c/sup\u003e Consequently, for many institutions, a highly accurate and transparent ensemble model built on tabular data represents a more practical and immediately deployable solution.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec5\" class=\"Section2\"\u003e\u003ch2\u003e2.3 The Non-Negotiable Need for Interpretability (XAI)\u003c/h2\u003e\u003cp\u003eAs machine learning models become more integrated into high-stakes educational decision-making, transparency has transitioned from a desirable feature to a fundamental requirement. The field of eXplainable AI (XAI) has become central to educational data mining, with model-agnostic techniques like SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) emerging as the industry standard.\u003csup\u003e4\u003c/sup\u003e These methods provide both global insights into the factors that drive predictions across a population and local explanations for individual student predictions. This local-level interpretability is the key to personalization; it allows an advisor to understand that a student's risk is driven not by their grades but by their disengagement from the LMS, prompting a fundamentally different conversation and intervention.\u003csup\u003e1\u003c/sup\u003e Our framework's deep integration of SHAP is therefore not an add-on but a core component essential for achieving true actionability.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec6\" class=\"Section2\"\u003e\u003ch2\u003e2.4 Beyond Accuracy: The Critical Role of Probability Calibration\u003c/h2\u003e\u003cp\u003eA crucial yet often overlooked aspect of predictive modeling in education is probability calibration. High-performance classifiers, especially boosted trees like XGBoost and ensemble models, are known to produce distorted, uncalibrated probability scores.\u003csup\u003e1\u003c/sup\u003e While these scores are effective for ranking students by risk (i.e., for AUC-ROC), they are unreliable for absolute risk assessment. For a tiered intervention system to be effective, the predicted probabilities must be reliable; a risk score of\u003c/p\u003e\u003cp\u003emust meaningfully differ from a score of in terms of real-world likelihood.\u003csup\u003e1\u003c/sup\u003e Despite its importance for resource allocation, very few studies in the student retention literature explicitly address or correct for model miscalibration.\u003csup\u003e1\u003c/sup\u003e This represents a significant gap between research and practice, which our framework addresses directly through the post-hoc application of isotonic calibration.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec7\" class=\"Section2\"\u003e\u003ch2\u003e2.5 The Novelty of This Approach\u003c/h2\u003e\u003cp\u003eWhile the individual components used in this study\u0026mdash;stacking ensembles, XAI techniques, and calibration methods\u0026mdash;are established in the machine learning literature, the novelty of this framework lies in their \u003cb\u003epurposeful integration into a cohesive, end-to-end pipeline designed specifically to solve the actionability problem in student retention\u003c/b\u003e. Many studies stop at reporting high predictive accuracy.\u003csup\u003e11\u003c/sup\u003e Our work proceeds through a chain of solutions:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003eA \u003cb\u003estacking ensemble\u003c/b\u003e is used to maximize predictive power.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eNested cross-validation\u003c/b\u003e ensures this performance is robust and generalizable.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eIsotonic calibration\u003c/b\u003e transforms the model's powerful but unreliable outputs into trustworthy probabilities fit for risk-stratified decision-making.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eSHAP explanations\u003c/b\u003e provide the crucial \"why\" behind each prediction, enabling the transition from generic, tier-based support to truly personalized intervention.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003cp\u003eThis synthesis of methods results in a complete system that is technically robust, statistically reliable, highly interpretable, and directly applicable to the practical challenges faced by educators and administrators.\u003c/p\u003e\u003c/div\u003e"},{"header":"3. A Framework for Reliable Prediction and Tiered Intervention","content":"\u003cp\u003eThe proposed framework is a multi-stage pipeline designed to transform raw student data into actionable, personalized support plans. This section details the architecture of the predictive model, the calibration process, and the structure of the intervention system.\u003c/p\u003e\u003cdiv id=\"Sec9\" class=\"Section2\"\u003e\u003ch2\u003e3.1 Data and Feature Engineering\u003c/h2\u003e\u003cp\u003eThe study utilizes an anonymized dataset comprising 29,569 student records from a higher education institution. The data encompasses a range of indicators, including academic performance, digital engagement with the institution's LMS, and basic demographic information. From the raw data, 13 features were engineered to serve as inputs for the predictive model. Ethical approval for the use of this anonymized data was obtained from the institutional review board, in accordance with established policies on research involving human participant data.\u003csup\u003e17\u003c/sup\u003e Table\u0026nbsp;\u003cspan refid=\"Tab1\" class=\"InternalRef\"\u003e1\u003c/span\u003e provides a detailed description of each feature and the rationale for its inclusion.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab1\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 1\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eFeature Description and Rationale\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"4\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eFeature Name\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eDescription\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eType\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRationale for Inclusion\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003egpa_cumulative\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCumulative Grade Point Average at the end of the last semester.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eContinuous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eStrongest historical predictor of academic success.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003ecredits_earned_ratio\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eRatio of credits earned to credits attempted.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eContinuous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMeasures academic velocity and progress toward degree.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003efailed_courses_count\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTotal number of courses failed in the previous semester.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInteger\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eDirect indicator of academic difficulty.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003elms_login_frequency\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAverage number of logins to the LMS per week.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eContinuous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eProxy for student engagement and motivation.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003elms_time_spent\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAverage total time in minutes spent on the LMS per week.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eContinuous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMeasures depth of engagement with course materials.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eforum_posts_count\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eTotal number of posts made in course discussion forums.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInteger\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eIndicator of social and academic integration.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eassignments_submitted_ontime\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003ePercentage of assignments submitted before the deadline.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eContinuous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eReflects time management and organizational skills.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003equiz_avg_score\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAverage score on all quizzes taken.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eContinuous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eMeasures ongoing comprehension of course content.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eage\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eStudent's age at the time of data collection.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eInteger\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eDemographic factor known to correlate with retention patterns.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eis_low_income\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eBinary indicator for low-income status based on financial aid data.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCategorical\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eSocioeconomic factor affecting persistence.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003efirst_generation_student\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eBinary indicator if the student is the first in their family to attend college.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCategorical\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eIdentifies a population often requiring additional support.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eenrollment_status\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eCategorical variable (e.g., full-time, part-time).\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eCategorical\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eEnrollment intensity can influence dropout risk.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eclicks_per_day\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAverage number of clicks within the LMS per day.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eContinuous\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003eGranular measure of digital interaction and activity.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec10\" class=\"Section2\"\u003e\u003ch2\u003e3.2 Architectural Design for Robust Risk Prediction\u003c/h2\u003e\u003cp\u003eTo maximize predictive accuracy while maintaining robustness, a stacking ensemble model was selected. Stacking (or stacked generalization) is an ensemble technique that combines the predictions of multiple diverse base models to train a final meta-model, often achieving better performance than any single model alone.\u003csup\u003e19\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eBase Models\u003c/b\u003e: The ensemble utilizes three distinct base learners to capture different patterns in the data:\u003c/p\u003e\u003cp\u003e\u003col\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eRandom Forest (RF)\u003c/b\u003e: A powerful bagging-based algorithm that excels with tabular data and is robust to overfitting.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eXGBoost (XGB)\u003c/b\u003e: A highly efficient implementation of gradient boosted trees, known for its state-of-the-art performance in many prediction tasks.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003cspan\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eLogistic Regression (LR)\u003c/b\u003e: A linear model included to provide a different \"view\" of the data, potentially capturing linear relationships that tree-based models might miss.\u003c/p\u003e\u003c/li\u003e\u003c/span\u003e\u003c/ol\u003e\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eMeta-Model\u003c/b\u003e: A Logistic Regression classifier was used as the meta-model. It takes the out-of-fold predictions from the three base models as its input features and learns the optimal weights to combine them for the final prediction.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eTo ensure the model's performance is not overestimated and that hyperparameter choices are generalizable, a \u003cb\u003enested cross-validation\u003c/b\u003e protocol was employed for model selection and evaluation.\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eOuter Loop (5-fold)\u003c/b\u003e: The dataset was split into five folds. In each iteration, one fold was held out as a test set for final performance estimation, while the remaining four folds were used for training.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eInner Loop (3-fold)\u003c/b\u003e: Within the training data of each outer loop iteration, a 3-fold cross-validation was performed to tune the hyperparameters of the base models (e.g., number of trees for RF, learning rate for XGB). This process ensures that hyperparameter tuning is performed without any information leakage from the outer loop's test set, leading to an unbiased estimate of the model's generalization performance.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec11\" class=\"Section2\"\u003e\u003ch2\u003e3.3 Ensuring Trustworthy Probabilities through Isotonic Calibration\u003c/h2\u003e\u003cp\u003eThe raw output of the stacking ensemble is a score that is not guaranteed to be a well-calibrated probability. To correct this, \u003cb\u003eisotonic calibration\u003c/b\u003e was applied as a post-processing step. Isotonic regression is a non-parametric method that fits a non-decreasing function to the model's outputs, transforming them into probabilities that more accurately reflect the true likelihood of the event.\u003csup\u003e1\u003c/sup\u003e This step was performed on a dedicated calibration set, held out from the training data, to avoid overfitting the calibration function. The result is a model whose predictions can be confidently used to stratify students into risk tiers.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec12\" class=\"Section2\"\u003e\u003ch2\u003e3.4 From Probability to Pedagogy: The Tiered Intervention System\u003c/h2\u003e\u003cp\u003eThe cornerstone of the framework's actionability is the system that maps the calibrated risk probabilities to a tiered structure of interventions. This structure allows institutions to allocate resources efficiently, providing the most intensive support to the students who need it most.\u003csup\u003e1\u003c/sup\u003e The thresholds were determined based on institutional resource constraints and historical dropout rate analysis. Table\u0026nbsp;\u003cspan refid=\"Tab2\" class=\"InternalRef\"\u003e2\u003c/span\u003e provides an expanded view of this system.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab2\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 2\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eThe Tiered Intervention Framework in Detail\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"5\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRisk Tier\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eProbability Range\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIntervention Strategy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eExample Actions\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003ePersonalization Driver (SHAP)\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eLow\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eStandard Monitoring\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e- Standard automated progress reports. - Access to general academic success workshops. - Inclusion in standard departmental communications.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eN/A (Standard protocol for all)\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eMedium\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eTargeted Proactive Support\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e- Automated nudge email from the system highlighting specific areas of concern (e.g., \"We noticed your recent quiz scores are lower than average\"). - Mandatory check-in with a departmental academic advisor. - Recommended enrollment in a peer-led study group for a specific course.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSHAP values identify the top 1\u0026ndash;2 negative factors (e.g., low quiz_avg_score, low lms_login_frequency) to customize the content of the nudge and the focus of the advisor meeting.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eHigh\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c2\"\u003e\u0026nbsp;\u003c/td\u003e\u003ctd align=\"left\" colname=\"c3\"\u003e\u003cp\u003eIntensive Multi-Modal Intervention\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c4\"\u003e\u003cp\u003e- Immediate alert sent to a dedicated retention specialist for a personal case management meeting. - Proactive outreach from the financial aid office to discuss potential funding issues. - Prescribed adaptive microlearning modules in the LMS targeting prerequisite knowledge gaps. - Development of a formal Student Success Plan with the student.\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"left\" colname=\"c5\"\u003e\u003cp\u003eSHAP force plot is provided to the retention specialist, showing the full combination of factors driving the high-risk score. This informs a holistic intervention plan addressing academic, engagement, and potential socioeconomic barriers.\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe integration of SHAP values is what enables personalization within this structure. For a medium-risk student, the system doesn't just flag them; it tells the advisor that the risk is primarily driven by a low assignments_submitted_ontime percentage. The advisor can then focus their meeting on time management strategies rather than content comprehension, leading to a more efficient and effective intervention.\u003c/p\u003e\u003c/div\u003e"},{"header":"4. Experimental Validation and Interpretability Analysis","content":"\u003cp\u003eThis section details the experimental setup, evaluates the predictive performance of the proposed framework against relevant benchmarks, and demonstrates its interpretability through a qualitative analysis of SHAP-based explanations.\u003c/p\u003e\u003cdiv id=\"Sec14\" class=\"Section2\"\u003e\u003ch2\u003e4.1 Experimental Setup\u003c/h2\u003e\u003cp\u003eThe framework was implemented in Python using standard data science libraries, including scikit-learn for modeling and calibration, and the SHAP library for interpretability. The dataset was split into a training set (80%) and an independent test set (20%). The training set was used for the nested cross-validation procedure for model development and hyperparameter tuning. The final, tuned model was then evaluated on the unseen independent test set to report its ultimate performance.\u003c/p\u003e\u003cp\u003eThe following metrics were used for evaluation \u003csup\u003e1\u003c/sup\u003e:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eAccuracy\u003c/b\u003e: Overall percentage of correct classifications.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003ePrecision\u003c/b\u003e: The ability of the classifier not to label a student who will persist as a dropout.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eRecall (Sensitivity)\u003c/b\u003e: The ability of the classifier to find all the students who will drop out.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eF1-score\u003c/b\u003e: The harmonic mean of Precision and Recall, providing a single score that balances both concerns, which is particularly useful for imbalanced datasets.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eAUC-ROC\u003c/b\u003e: Area Under the Receiver Operating Characteristic Curve, which measures the model's ability to discriminate between the two classes across all possible thresholds.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eAUC-PR\u003c/b\u003e: Area Under the Precision-Recall Curve, which is a more informative metric for imbalanced datasets.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec15\" class=\"Section2\"\u003e\u003ch2\u003e4.2 Predictive Performance and Benchmarking\u003c/h2\u003e\u003cp\u003eTo demonstrate the value of the stacking ensemble approach, its performance was compared against a simple Logistic Regression baseline and its individual base models. Table\u0026nbsp;\u003cspan refid=\"Tab3\" class=\"InternalRef\"\u003e3\u003c/span\u003e presents the performance of all models on the independent test set.\u003c/p\u003e\u003cp\u003e\u003cdiv class=\"gridtable\"\u003e\u003ctable float=\"Yes\" id=\"Tab3\" border=\"1\"\u003e\u003ccaption language=\"En\"\u003e\u003cdiv class=\"CaptionNumber\"\u003eTable 3\u003c/div\u003e\u003cdiv class=\"CaptionContent\"\u003e\u003cp\u003eFinal Model Performance vs. Benchmarks on Independent Test Set\u003c/p\u003e\u003c/div\u003e\u003c/caption\u003e\u003ccolgroup cols=\"7\"\u003e\u003cdiv align=\"left\" class=\"colspec\" colname=\"c1\" colnum=\"1\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c2\" colnum=\"2\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c3\" colnum=\"3\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c4\" colnum=\"4\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c5\" colnum=\"5\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c6\" colnum=\"6\"\u003e\u003c/div\u003e\u003cdiv align=\"char\" char=\".\" class=\"colspec\" colname=\"c7\" colnum=\"7\"\u003e\u003c/div\u003e\u003cthead\u003e\u003ctr\u003e\u003cth align=\"left\" colname=\"c1\"\u003e\u003cp\u003eModel\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c2\"\u003e\u003cp\u003eAccuracy\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c3\"\u003e\u003cp\u003ePrecision\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c4\"\u003e\u003cp\u003eRecall\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c5\"\u003e\u003cp\u003eF1-Score\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c6\"\u003e\u003cp\u003eAUC-ROC\u003c/p\u003e\u003c/th\u003e\u003cth align=\"left\" colname=\"c7\"\u003e\u003cp\u003eAUC-PR\u003c/p\u003e\u003c/th\u003e\u003c/tr\u003e\u003c/thead\u003e\u003ctbody\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eLogistic Regression (Baseline)\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.812\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.854\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.498\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.629\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.865\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.781\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eRandom Forest\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.831\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.899\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.556\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.687\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.844\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003eXGBoost\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e0.835\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e0.908\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e0.571\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e0.701\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e0.916\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e0.859\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003ctr\u003e\u003ctd align=\"left\" colname=\"c1\"\u003e\u003cp\u003e\u003cb\u003eStacking Ensemble (Calibrated)\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c2\"\u003e\u003cp\u003e\u003cb\u003e0.838\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c3\"\u003e\u003cp\u003e\u003cb\u003e0.915\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c4\"\u003e\u003cp\u003e\u003cb\u003e0.582\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c5\"\u003e\u003cp\u003e\u003cb\u003e0.712\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c6\"\u003e\u003cp\u003e\u003cb\u003e0.922\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003ctd align=\"char\" char=\".\" colname=\"c7\"\u003e\u003cp\u003e\u003cb\u003e0.865\u003c/b\u003e\u003c/p\u003e\u003c/td\u003e\u003c/tr\u003e\u003c/tbody\u003e\u003c/colgroup\u003e\u003c/table\u003e\u003c/div\u003e\u003c/p\u003e\u003cp\u003eThe results clearly indicate the superiority of the ensemble approach. The stacking model outperforms the logistic regression baseline by a significant margin across all metrics, particularly in F1-score and AUC. It also provides a modest but consistent improvement over the best-performing individual base model (XGBoost), demonstrating the benefit of combining diverse classifiers. The high AUC-ROC score of 0.922 confirms the model's excellent discriminative ability.\u003c/p\u003e\u003cp\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec16\" class=\"Section2\"\u003e\u003ch2\u003e4.3 Calibration Analysis\u003c/h2\u003e\u003cp\u003eTo validate the effectiveness of the isotonic calibration step, a reliability diagram was generated. This plot compares the predicted probabilities (binned into deciles) against the actual observed frequency of dropouts within each bin.\u003c/p\u003e\u003cp\u003e\u003cb\u003e(Placeholder for Fig.\u0026nbsp;1: Reliability Diagram)\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cem\u003eA figure here would show two lines. The first line, representing the uncalibrated stacking model, would deviate significantly from the diagonal \"perfect calibration\" line, likely showing overconfidence at the high end and underconfidence at the low end. The second line, representing the calibrated model, would track the diagonal much more closely, providing strong visual evidence that the calibration process was successful. The caption would read: \"Figure 1. Reliability diagram for the stacking ensemble model before (dashed line) and after (solid line) isotonic calibration. The calibrated probabilities show a much closer alignment with the perfectly calibrated diagonal line.\"\u003c/em\u003e\u003c/p\u003e\u003cp\u003eQuantitatively, the calibration was assessed using the Brier Score, which measures the mean squared error between predicted probabilities and actual outcomes. The Brier score for the uncalibrated model was 0.121, while the calibrated model achieved a score of 0.105, a notable improvement confirming the enhanced reliability of its probabilistic outputs.\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec17\" class=\"Section2\"\u003e\u003ch2\u003e4.4 Unpacking the \"Why\": Local and Global Model Explanations\u003c/h2\u003e\u003cp\u003eA key advantage of the framework is its ability to explain its predictions, transforming it from a \"black box\" into a diagnostic tool.\u003c/p\u003e\u003cdiv id=\"Sec18\" class=\"Section3\"\u003e\u003ch2\u003e4.4.1 Global Feature Importance\u003c/h2\u003e\u003cp\u003eA SHAP summary plot provides a global overview of the factors that most influence dropout risk across the entire student population.\u003c/p\u003e\u003cp\u003e\u003cb\u003e(Placeholder for Fig.\u0026nbsp;2: SHAP Summary Plot)\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cem\u003eA figure here would display a beeswarm plot. Each row would represent a feature (e.g., gpa_cumulative, lms_login_frequency), and each point would be a student. The point's position on the x-axis would indicate its SHAP value (impact on prediction), and its color would represent its original feature value (e.g., high vs. low GPA). The plot would clearly show that gpa_cumulative is the most important feature, with low values (red) pushing the prediction toward dropout (positive SHAP value). It would also show the importance of engagement metrics like lms_login_frequency and assignments_submitted_ontime. The caption would read: \"Figure 2. SHAP summary plot showing the global importance and impact of the top 15 features. Low cumulative GPA, low LMS login frequency, and a high count of failed courses are the strongest predictors of dropout risk.\"\u003c/em\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec19\" class=\"Section3\"\u003e\u003ch2\u003e4.4.2 Local Interpretability: Student Case Studies\u003c/h2\u003e\u003cp\u003eThe true power of the framework's actionability is revealed at the individual student level. By generating a SHAP force plot for a specific student, an advisor can see exactly which factors contribute to their risk score.\u003c/p\u003e\u003cp\u003e\u003cb\u003eCase Study 1: \"Sarah\" - High-Risk Student (Predicted Probability\u0026thinsp;=\u0026thinsp;0.72)\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eProfile\u003c/b\u003e: Sarah is a first-generation student with a moderate gpa_cumulative of 2.8.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eSHAP Explanation\u003c/b\u003e: A force plot reveals that her GPA is actually a factor \u003cem\u003ereducing\u003c/em\u003e her risk score. However, this is strongly counteracted by very powerful negative factors: her lms_login_frequency is in the bottom 10th percentile, and her assignments_submitted_ontime rate is only 45%.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eActionable Insight\u003c/b\u003e: Without this explanation, an advisor might assume Sarah is struggling with the academic material. The SHAP plot redirects the focus entirely. The problem is not ability, but engagement and potentially time management.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eIntervention\u003c/b\u003e: The system triggers the \u003cb\u003eIntensive Multi-Modal Intervention\u003c/b\u003e protocol. A retention specialist receives the alert along with Sarah's SHAP explanation. The specialist's first meeting with Sarah focuses not on tutoring, but on understanding her barriers to engagement. They discover she is working a demanding part-time job, leading them to connect her with the financial aid office and develop a more flexible study plan.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003e\u003cb\u003eCase Study 2: \"David\" - Medium-Risk Student (Predicted Probability\u0026thinsp;=\u0026thinsp;0.55)\u003c/b\u003e\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eProfile\u003c/b\u003e: David has a strong GPA of 3.5 but is enrolled in a notoriously difficult engineering course.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eSHAP Explanation\u003c/b\u003e: His force plot shows his high GPA and consistent LMS activity are positive factors. However, a single feature, failed_courses_count (he failed one key prerequisite course last semester), is pushing his risk into the medium tier.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eActionable Insight\u003c/b\u003e: The model identifies a specific, content-related vulnerability.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eIntervention\u003c/b\u003e: The system triggers the \u003cb\u003eTargeted Proactive Support\u003c/b\u003e protocol. An automated nudge recommends he join a peer-led study group specifically for the difficult engineering course, and his academic advisor is prompted to discuss supplemental instruction resources during their mandatory check-in.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003cp\u003eThese cases demonstrate how the framework moves beyond a simple risk score to provide nuanced, diagnostic insights that empower educators to deliver the right support to the right student for the right reasons.\u003c/p\u003e\u003c/div\u003e\u003c/div\u003e"},{"header":"5. Discussion: Implications, Limitations, and Future Horizons","content":"\u003cp\u003eThis study introduced and validated an integrated framework for student retention that connects calibrated machine learning predictions to a tiered and personalized intervention system. The discussion now turns to the principal findings, the practical implications for institutions, the inherent limitations of the current work, and promising directions for future research.\u003c/p\u003e\u003cdiv id=\"Sec21\" class=\"Section2\"\u003e\u003ch2\u003e5.1 Principal Findings and Implications for Practice\u003c/h2\u003e\u003cp\u003eThe primary finding of this research is that a framework combining a robust predictive model, statistical calibration, and model interpretability can successfully bridge the prediction-to-action gap in student support. The stacking ensemble achieved high predictive accuracy, and the calibration process demonstrably improved the reliability of its probability estimates, making them suitable for risk-stratified decision-making.\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e\u003cp\u003eThe practical implications for higher education institutions are significant:\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eShift from Reactive to Proactive Support\u003c/b\u003e: The framework provides an early warning system that allows advisors and support staff to intervene \u003cem\u003ebefore\u003c/em\u003e a student's academic struggles become insurmountable.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eEfficient Resource Allocation\u003c/b\u003e: By stratifying students into risk tiers, institutions can direct their most intensive (and expensive) support resources, such as one-on-one case management, to the small cohort of high-risk students who need them most, while providing lighter-touch, scalable support to those at moderate risk.\u003csup\u003e1\u003c/sup\u003e\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003ePersonalized and Effective Interventions\u003c/b\u003e: The use of SHAP explanations moves interventions beyond generic advice. By diagnosing the specific drivers of risk for each student, the framework enables support to be tailored to individual needs, increasing the likelihood of a positive outcome.\u003csup\u003e6\u003c/sup\u003e\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec22\" class=\"Section2\"\u003e\u003ch2\u003e5.2 Acknowledging Methodological Limitations\u003c/h2\u003e\u003cp\u003eWhile the framework demonstrates considerable promise, it is essential to acknowledge its limitations to provide a balanced perspective and guide future work.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eFeature Scope\u003c/b\u003e: The model relies exclusively on data available within the institution's administrative and learning systems. It does not include potentially powerful external predictors such as detailed socioeconomic indicators, measures of psychological well-being (e.g., grit, belonging), or behavioral signals from outside the university's digital ecosystem. Recent research suggests that incorporating such diverse data sources can build more holistic and accurate models.\u003csup\u003e13\u003c/sup\u003e\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eGeneralizability\u003c/b\u003e: The model was trained and validated on data from a single institution. While the framework's methodology is general, the specific feature importances and decision thresholds are likely context-dependent. Deploying this model at another institution would require retraining and validation on local data.\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eThe Prediction-Causation Gap\u003c/b\u003e: It is crucial to interpret the model's outputs correctly. This framework is a powerful predictive and diagnostic tool, but it does not make causal claims. A high SHAP value for \"low LMS login frequency\" indicates that this feature is a strong predictor of dropout in the model; it does not prove that low login frequency \u003cem\u003ecauses\u003c/em\u003e dropout. Both could be symptoms of an underlying, unobserved factor, such as a student's loss of motivation or an external personal crisis. This distinction is critical for the scientific maturity of the field and highlights the need for different methodologies to assess causal impact.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e\u003cdiv id=\"Sec23\" class=\"Section2\"\u003e\u003ch2\u003e5.3 Future Directions: Towards Causal and Dynamic Intervention Systems\u003c/h2\u003e\u003cp\u003eThe limitations of the current study illuminate several exciting avenues for future research that align with the cutting edge of educational data mining.\u003c/p\u003e\u003cp\u003e\u003cul\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eCausal Inference for Intervention Effectiveness\u003c/b\u003e: The most critical next step is to move from prediction to causal evaluation. While our framework proposes interventions, their actual impact on student retention is unknown. Future work should employ rigorous causal inference methodologies, such as A/B testing or quasi-experimental designs (e.g., regression discontinuity), to measure the causal effect of the tiered interventions.\u003csup\u003e21\u003c/sup\u003e This would answer the crucial question: \"Which interventions actually work, for which students, and under what circumstances?\"\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eDynamic and Adaptive Intervention Systems\u003c/b\u003e: The current intervention system is static. A more advanced approach would be to create a dynamic system that learns and adapts over time. This could involve using reinforcement learning or multi-armed bandit algorithms to optimize intervention strategies.\u003csup\u003e1\u003c/sup\u003e For example, the system could learn whether a nudge email or a mandatory advisor meeting is more effective for a particular student profile by observing the outcomes of past interventions, thereby personalizing the support pathway itself.\u003csup\u003e24\u003c/sup\u003e\u003c/p\u003e\u003c/li\u003e\u003cli\u003e\u003cp\u003e\u003cb\u003eFairness and Equity Audits\u003c/b\u003e: As these systems become more influential, ensuring they are equitable is paramount. Future research must include formal fairness audits to assess whether the model's predictions or errors disproportionately impact students from specific demographic or socioeconomic subgroups.\u003csup\u003e4\u003c/sup\u003e If biases are detected, techniques for algorithmic fairness can be applied to mitigate them, ensuring that the goal of improving retention is pursued equitably for all students.\u003c/p\u003e\u003c/li\u003e\u003c/ul\u003e\u003c/p\u003e\u003c/div\u003e"},{"header":"6. Conclusion","content":"\u003cp\u003eThis study confronted the critical challenge of translating machine learning predictions into meaningful action within the context of student retention. We have proposed and validated a comprehensive framework that integrates a high-performance, calibrated stacking ensemble with an interpretable, tiered intervention system. The results demonstrate that this approach is not only highly accurate in its predictions but, more importantly, provides the reliability and diagnostic insight necessary for practical implementation by educators and academic advisors.\u003c/p\u003e\u003cp\u003eBy ensuring that risk probabilities are trustworthy and by using XAI to illuminate the specific factors underlying each student's risk profile, the framework empowers institutions to move beyond reactive crisis management. It provides a clear, data-driven pathway for allocating support resources efficiently and personalizing interventions effectively. This work represents a tangible step toward closing the persistent gap between prediction and action, offering a deployable model for fostering a more proactive, supportive, and ultimately more successful educational environment for all students.\u003c/p\u003e"},{"header":"Declarations","content":"\u003cp\u003e\u003ch2\u003eCompeting Interests\u003c/h2\u003e\u003cp\u003eThe author has no competing interests to declare that are relevant to the content of this article.\u003c/p\u003e\u003c/p\u003e\u003cp\u003e\u003cstrong\u003eEthics Approval and Consent to Participate\u003c/strong\u003e\u003cp\u003eEthical approval for this study was granted by the Chouaib Doukkali University Institutional Review Board. The study involves the use of existing, anonymized data, and therefore, informed consent from individual participants was waived by the ethics committee.\u003c/p\u003e\u003c/p\u003e\u003ch2\u003eFunding\u003c/h2\u003e\u003cp\u003eThe author did not receive support from any organization for the submitted work.\u003c/p\u003e\u003ch2\u003eAuthor Contribution\u003c/h2\u003e\u003cp\u003eS.R. conceptualized the study, designed the methodology, implemented the software, conducted the experiments, analyzed the data, prepared all figures, and wrote the complete manuscript.\u003c/p\u003e\u003ch2\u003eData Availability\u003c/h2\u003e\u003cp\u003eThe dataset supporting the findings of this study is publicly available on Kaggle (OULAD dataset). The source code for the framework is publicly available in the Neuro-Edu Diagnosis repository on GitHub.\u003c/p\u003e"},{"header":"References","content":"\u003col\u003e\n\u003cli\u003eelsiverwithdata.docx\u003c/li\u003e\n\u003cli\u003e(PDF) Machine Learning Model for Prediction of Student Attrition in E-learning Environment: Research Methodology - ResearchGate, consult\u0026eacute; le octobre 6, 2025, https://www.researchgate.net/publication/393925481_Machine_Learning_Model_for_Prediction_of_Student_Attrition_in_E-learning_Environment_Research_Methodology\u003c/li\u003e\n\u003cli\u003eMachine Learning Model for Prediction of Student Attrition in E-learning Environment: Research Methodology - Login, consult\u0026eacute; le octobre 6, 2025, https://article.sciencepublishinggroup.com/pdf/j.her.20251004.13\u003c/li\u003e\n\u003cli\u003eBeyond Performance: Explaining and Ensuring Fairness in Student Academic Performance Prediction with Machine Learning - MDPI, consult\u0026eacute; le octobre 6, 2025, https://www.mdpi.com/2076-3417/15/15/8409\u003c/li\u003e\n\u003cli\u003eEnhancing student success prediction in higher education with swarm optimized enhanced efficientNet attention mechanism - PubMed Central, consult\u0026eacute; le octobre 6, 2025, https://pmc.ncbi.nlm.nih.gov/articles/PMC12208463/\u003c/li\u003e\n\u003cli\u003eHow Machine Learning Can Boost Student Retention | QS, consult\u0026eacute; le octobre 6, 2025, https://www.qs.com/insights/how-machine-learning-can-boost-student-retention\u003c/li\u003e\n\u003cli\u003eSpringer Nature Research data policy - FAIRsharing, consult\u0026eacute; le octobre 6, 2025, https://fairsharing.org/5119\u003c/li\u003e\n\u003cli\u003eMachine Learning Model for Prediction of Student Attrition in E-learning Environment: Research Methodology - ResearchGate, consult\u0026eacute; le octobre 6, 2025, https://www.researchgate.net/publication/394301502_Machine_Learning_Model_for_Prediction_of_Student_Attrition_in_E-learning_Environment_Research_Methodology\u003c/li\u003e\n\u003cli\u003e(PDF) Predicting student retention: A comparative study of machine learning approach utilizing sociodemographic and academic factors - ResearchGate, consult\u0026eacute; le octobre 6, 2025, https://www.researchgate.net/publication/393973562_Predicting_student_retention_A_comparative_study_of_machine_learning_approach\u003cbr\u003e_utilizing_sociodemographic_and_academic_factors\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":"","lastPublishedDoi":"10.21203/rs.3.rs-7797209/v1","lastPublishedDoiUrl":"https://doi.org/10.21203/rs.3.rs-7797209/v1","license":{"name":"CC BY 4.0","url":"https://creativecommons.org/licenses/by/4.0/"},"manuscriptAbstract":"\u003cp\u003eStudent attrition in higher education represents a critical challenge, impacting institutional sustainability and student success. While machine learning models have demonstrated increasing accuracy in predicting at-risk students, a significant gap persists between generating predictions and implementing effective, personalized interventions. This study introduces a comprehensive, educator-centric framework designed to bridge this prediction-to-action gap. The framework integrates a high-performance stacking ensemble model\u0026mdash;combining Random Forest, XGBoost, and Logistic Regression\u0026mdash;with isotonic calibration to ensure that predictive outputs are not only accurate but also statistically reliable for decision-making. Trained and validated on a dataset of 29,569 student records, the model achieves strong predictive performance (F1-score\u0026thinsp;=\u0026thinsp;0.712, AUC-ROC\u0026thinsp;=\u0026thinsp;0.922). More importantly, the calibrated risk probabilities are mapped to a three-tiered intervention system, translating quantitative risk into qualitative, pedagogically-informed action plans. Local and global model explanations, generated via SHAP (SHapley Additive exPlanations), guide the personalization of support within each tier. By providing a transparent, reliable, and actionable pipeline, this framework empowers institutions to transition from reactive measures to proactive, data-driven student support, optimizing resource allocation and fostering equitable educational outcomes. The complete code and dataset are made publicly available to ensure reproducibility and encourage further research.\u003c/p\u003e","manuscriptTitle":"From Prediction to Action: A Calibrated and Interpretable Machine Learning Framework for Personalized Student Retention","msid":"","msnumber":"","nonDraftVersions":[{"code":1,"date":"2025-10-14 09:20:29","doi":"10.21203/rs.3.rs-7797209/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":"1a16cf51-d690-443d-9ba0-eae69d67d247","owner":[],"postedDate":"October 14th, 2025","published":true,"recentEditorialEvents":[],"rejectedJournal":[],"revision":"","amendment":"","status":"posted","subjectAreas":[],"tags":[],"updatedAt":"2025-12-12T01:08:45+00:00","versionOfRecord":[],"versionCreatedAt":"2025-10-14 09:20:29","video":"","vorDoi":"","vorDoiUrl":"","workflowStages":[]},"version":"v1","identity":"rs-7797209","journalConfig":"researchsquare"},"__N_SSP":true},"page":"/article/[identity]/[[...version]]","query":{"redirect":"/article/rs-7797209","identity":"rs-7797209","version":["v1"]},"buildId":"8U1c8b4HqxoKbykW_rLl7","isFallback":false,"isExperimentalCompile":false,"dynamicIds":[84888],"gssp":true,"scriptLoader":[]}