Cross-Institution Student Performance Prediction Using an Explainable Stacking Ensemble Framework
TL;DR
The experimental results demonstrated that the proposed stacking model achieves strong predictive performance, with an accuracy of 95.82% and an Area Under the ROC Curve (AUC) close to 1.0 on the primary dataset, while maintaining strong generalization across external datasets.
Abstract
This study presents an explainable machine learning framework for early prediction of student academic performance using a stacking-based ensemble approach. The proposed model integrates Random Forest (RF), XGBoost, and Support Vector Machine (SVM) as base learners, with Logistic Regression (LR) used as a meta-learner. A dataset of 2,392 student records, containing demographic and behavioral features, was used for training and evaluation. Data preprocessing included encoding and normalization, followed by model optimization using cross-validation and grid search. The experimental results demonstrated that the proposed stacking model achieves strong predictive performance, with an accuracy of 95.82% and an Area Under the ROC Curve (AUC) close to 1.0 on the primary dataset, while maintaining strong generalization across external datasets. Model interpretability is enhanced using SHapley Additive exPlanations (SHAP) analysis, which identifies Grade Point Average (GPA), attendance, and study time as the most influential factors. In addition, the framework is implemented as a web-based system for real-time prediction and decision support. The results demonstrate that the proposed approach is a practical solution for educational data analytics.