Multimodal University Academic Performance Prediction Model Based on Heterogeneous Graph Neural Networks

Abstract

University academic performance prediction faces three major challenges: difficulty in semantic alignment of multisource heterogeneous data, sparse graph structure, and modality imbalance. To address this, this paper proposes a Heterogeneous Graph Network for Multimodal Student Performance Prediction (HGN‐MSP) model. First, the model integrates the three modalities of grades, behavior, and resources to construct a campus knowledge graph. Second, a meta‐path‐guided feature aggregation mechanism is designed to achieve cross‐modal semantic alignment. Dynamic Neighbor Sampling (DNS) is applied to alleviate the lack of information aggregation caused by graph sparsity. Adversarial Modality Balancing (AMB) is proposed to suppress the dominance of a single modality and enhance model robustness. Finally, the Shapley Additive exPlanations (SHAP) framework is integrated to improve interpretability. On a self‐built dataset of 19,856 students, the HGN‐MSP model achieves an Area under the ROC Curve (AUC) of 0.896, a significant 5.3% improvement over the baseline model, eXtreme Gradient Boosting (XGBoost), and also achieves top performance in F1‐score and Recall, demonstrating its effectiveness and superiority in precisely identifying students at academic risk.