An Ensemble Learning Approach with Explainability for Early Prediction of Student Academic Performance

Abstract

This paper proposes an interpretable ensemble learning framework that integrates multi-source student features including demographic attributes, family background, behavioral indicators, and prior academic records to predict final academic performance using the
widely adopted UCI Student Performance dataset. Five machine learning classifiers are systematically compared: Logistic Regression, Decision Tree, Random Forest, XGBoost, and LightGBM. To address class imbalance, the Synthetic Minority Oversampling Technique (SMOTE)
is applied during preprocessing. Experimental results demonstrate that XGBoost achieves the best overall performance, with an accuracy of
87.3%, F1-score of 0.851, and AUC-ROC of 0.923. Furthermore, SHapley Additive exPlanations (SHAP) are employed to provide featurelevel interpretability, revealing that prior grades, study time, and number of past course failures are the most influential predictors. The proposed framework not only advances predictive accuracy but also offers actionable insights for educators and policymakers seeking to implement data-driven early warning systems.