A semantic-aligned stacked ensemble framework for cross-semester student performance prediction

Abstract

Accurate prediction of student academic performance is crucial for the advancement of intelligent educational systems. However, significant challenges persist in achieving high prediction accuracy and strong generalization capability, particularly in data-scarce educational scenarios with heterogeneous assessment structures across semesters. This paper presents an empirically motivated semantic-aligned stacked ensemble framework that addresses the limitations of low prediction accuracy and poor generalization in data-scarce student performance prediction. The framework combines three targeted design elements: 13 temporally grounded statistical features that establish a semantically comparable feature space across heterogeneous assessment structures; a two-stage adaptation pipeline which is CORAL covariance alignment followed by MMD-guided per-feature weighting that narrows cross-domain distribution discrepancies; and a systematically designed heterogeneous ensemble with Ridge-regularized stacking that ensures robust aggregation in small-sample regimes. Extensive experiments on a real-world cross-semester dataset demonstrate that the proposed method outperforms state-of-the-art baselines, achieving an R2 of 0.5425 and an MAE of 0.0851 on the target domain test set.