An Explainable Artificial Intelligence for Student Academic Performance Prediction: A Comparative Analysis of Different Models and Explanation Techniques

Student academic performance prediction enables early identification of at-risk learners, but predictive accuracy alone cannot ensure transparent and reliable educational decisions. This review comparatively analyzes recent Explainable Artificial Intelligence (XAI) approaches across datasets, prediction tasks, models, validation strategies, and explanation methods. SHAP is the dominant explanation technique, frequently combined with tree-based models, while LIME provides complementary local explanations. Among classification studies, LSTM-SHAP achieved 92.88% accuracy and 92.79% F1-score on OULAD, while Random Forest achieved 90.77% accuracy for academic-risk prediction. For regression, TabPFN-SHAP reported an R2 of 0.9216. However, these results cannot be directly ranked because datasets, targets, and validation protocols differ. Overall, tree-based XAI models provide a strong accuracy–interpretability balance, while LSTM-SHAP and TabPFN-SHAP show advantages for sequential and tabular educational data, respectively. The review identifies external validation, fairness, explanation stability, and actionable intervention as key directions for future research.