Towards explainable knowledge tracing through meta-path-free heterogeneous graph learning with large language models

Knowledge tracing (KT) is fundamental to intelligent tutoring systems because it models student knowledge evolution and predicts future learning performance. However, existing approaches often struggle to simultaneously capture heterogeneous educational relationships, long-term temporal dependencies, and semantic information embedded in instructional content, limiting predictive accuracy, explainability, and cross-dataset transfer. To address these challenges, this paper proposes MHG-KT (Meta-path-free Heterogeneous Graph Knowledge Tracing), an explainable framework that integrates a meta-path-free heterogeneous graph encoder, Transformer-based temporal modeling, and pretrained Llama 3 8B semantic embeddings through a joint cross-modal attention mechanism. Unlike conventional approaches that learn graph and semantic representations independently before feature fusion, MHG-KT enables continuous interaction between heterogeneous relational and semantic representations during temporal learning, allowing structural, temporal, and semantic dependencies to be jointly optimized within a unified framework. The heterogeneous graph encoder models relationships among students, skills, problems, and contextual features without manually engineered meta-paths, while the Transformer captures knowledge evolution and the semantic encoder enriches graph representations using problem statements and instructional hints. The proposed framework was evaluated on three benchmark datasets (ASSISTments, EdNet, and Junyi) against four representative knowledge tracing models (DIMKT, simpleKT, GIKT, and TGNN). Experimental results demonstrate that MHG-KT consistently outperformed all baselines, achieving AUC scores of 0.927, 0.922, and 0.934 on ASSISTments, EdNet, and Junyi, respectively. Compared with the strongest baseline, the proposed framework improved AUC by up to 1.7%, achieved prediction accuracy of 89.0%, reduced trajectory forecasting error to a minimum MAE of 0.033, maintained calibration errors below 0.061 across all datasets, and achieved inference latency below 5 ms per student interaction.