Temporal Knowledge Enhanced Graph Attention Network for Explainable Learning Path Recommendation

ABSTRACT

To address the issues of weak dynamic interactivity and insufficient interpretability in personalized learning path recommendation, this paper proposes a Temporal Knowledge‐Enhanced Graph Attention Network (TKEGAT). First, based on questionnaire statistical analysis, this study quantifies learners' path preferences across four typical learning scenarios: initial learning, daily review, pre‐examination learning, and pre‐examination review, and verifies the influence patterns of learning ability and learning scenarios on such preferences. Second, a recommendation model covering the complete learning cycle is constructed. It integrates course nodes, the Ebbinghaus forgetting curve incorporating learning ability factors, and a review mechanism to accurately simulate the dynamic evolution of learners' knowledge states. Third, a four‐dimension weighted interpretable reward function is designed. It comprehensively considers preference matching, structural relevance, temporal rationality, and model confidence, unifying and quantifying personalized preferences, knowledge graph structural constraints, course progression logic, and model prediction confidence into instant feedback signals to drive the Deep Q ‐Network for sequential decision optimization. Meanwhile, the Graph Attention Network is used to mine structural features of the knowledge graph for enhanced state representation. Finally, using experimental data from real course learners, the effectiveness of the model is verified through ablation experiments and baseline comparisons. TKEGAT significantly outperforms all comparative methods in explainability scores while maintaining high path accuracy, demonstrating that it can effectively generate personalized learning paths with both rationality and interpretability.