An Explainable Educational Prediction Model Based on a Hierarchical Temporal Graph Network

Abstract

Educational prediction aims to estimate future learning outcomes based on students’ historical records, course performance, and behavioral features. This task is challenging because learning processes contain both long-term trends and short-term fluctuations, educational data involve hierarchical relations among students, courses, and cohorts, and many existing models provide limited explanation for their predictions. To address these issues, this study proposes a Hierarchical Temporal Graph Network for Educational Prediction (HTGN–EP). The model integrates long–short temporal modeling, hierarchical graph representation, and multi-task optimization into a unified framework to capture the temporal evolution of learning behaviors and the relational structure among students, courses, and cohorts. Experiments on real educational datasets show that HTGN–EP outperforms existing baseline methods, achieving an MAE of 0.136, an RMSE of 0.171, and an $$R^{2}$$ of 0.892, while maintaining stable performance across multiple prediction tasks. Further gradient-based attribution analysis indicates that prior GPA, attendance rate, and study time have strong effects on prediction outcomes, suggesting that the model can provide useful explanatory information while improving predictive accuracy. Overall, HTGN–EP effectively models temporal changes and hierarchical dependencies in educational data, offering an interpretable approach for learning outcome prediction, student risk identification, and educational management decision making.

The authors' abstract, as published at the source. Neural Processing Letters, 2026 · DOI ↗