Explainable machine learning with class balancing for predicting student academic performance

Abstract

Accurate prediction of student academic performance is crucial for implementing effective early intervention strategies in higher education. However, existing models often struggle with class imbalance and lack interpretability, limiting their practical adoption. This methodological proof-of-concept develops a robust predictiveframework by benchmarking multiple machine learning algorithms, specifically Logistic Regression, Decision Tree, and the standard Random Forest, against a proposed Random Forest classifier with class balancing. Utilizing a public simulation dataset of 2,392 student records, the proposed model incorporates algorithmic penalties for misclassifications of the minority class to address data skewness. Experimental results demonstrate that this balanced approach significantly outperforms the linear and standard tree-based baselines, achieving an Accuracy of 92.3% and an F1 Macro score of 0.872. The results confirm the model's superior capability in identifying at-risk students without compromising overall precision. Furthermore, explainability analysis using SHAP (Shapley Additive exPlanations) identifies Grade Point Average (GPA) as the dominant predictor, while highlighting Absences and Weekly Study Time as critical behavioral leading indicators. These findings demonstrate that integrating predictive accuracy with explainable AI supports proactive and evidence-based academic governance. Future research is recommended to validate this framework using longitudinal data streams within Learning Management Systems, enabling real-time monitoring.