Explainable Hierarchical Multi-Task Learning for Multi-Level Student Performance Prediction

Academic achievement prediction by students has been proven to be one of the essential problems in educational data mining, which has immediate consequences on preventive measures and future planning within any educational institution. The methods proposed to solve this problem until now have focused on solving individual prediction problems independently without being able to account for the hierarchical interdependencies of grades obtained from assignments, classes, and overall semester academic performance. This work introduces a novel framework called X-HMTL-SP for the aforementioned purposes. The framework integrates a Shared Feature Encoder, three hierarchically chained Task-Specific Prediction Heads (Random Forest, Gradient Boosting Classifier, and RF Regressor), and a Permutation Feature Importance (PFI) Explainability Module. X-HMTL-SP has been tested using an integrated database comprising 5,468 students' data across three heterogeneous datasets from the UCI education repository and shows 99.82% precision for predicting assignments, 100% precision for course grades (regression R² = 0.9993), and 81.26% precision for semester performance predictions. The cross-validation results show consistent generalization ability (Tasks 1 and 3 CV F1 scores are 0.9979 ±. Explain ability analysis demonstrates that hierarchical signals from lower-level heads rank among the top-6 most important features for semester-level prediction, directly validating the knowledge transfer mechanism. The proposed framework advances multi-level educational prediction with actionable, interpretable insights for academic early-warning systems.