Predicting Student Academic Performance Based on Multi-source Heterogeneous Features and Explainable Ensemble Learning

Abstract

Student academic performance prediction is a key task in learning analysis and educational data mining, aiming to achieve early risk identification, precise intervention, and teaching decision support through multi-source educational data. Addressing the issues of insufficient validation of public data and real-world scenarios, inadequate evaluation protocols, and weak stability of interpretation results in existing research, this paper proposes a multi-source heterogeneous feature modeling and interpretable ensemble learning framework for educational intervention. This framework integrates demographic features, previous academic foundation, behavioral participation, psychological state, and teaching context information, and systematically compares linear models, random forests, boosting models, and fusion models in a unified leak-proof modeling pipeline. The model performance is evaluated through repeated cross-validation, ablation experiments, missing robustness analysis, and statistical significance tests. At the interpretation level, this paper combines permutation importance and SHAP to analyze key influencing factors from the global, local, and stability perspectives. Experimental results show that the boosting models and fusion models overall outperform the linear baseline in terms of accuracy and robustness; previous academic performance, attendance participation, homework completion rate, and behavioral investment are the most stable core predictors, while psychological stress and teaching quality variables provide important supplementary explanations. This paper has formed a reproducible, scalable, and education-intervention-oriented academic performance prediction technology route, providing methodological basis for university academic warning and personalized support.