Cluster-enhanced behavioral knowledge tracing for interpretable knowledge evolution

Abstract

Knowledge tracing aims to infer students’ latent knowledge states from historical interaction sequences and serves as a fundamental technique for personalized learning and intelligent assessment. However, existing methods predominantly rely on sequence-level modeling and fail to adequately capture the complex interplay among behavioral patterns, content structures, and dynamic forgetting. As a result, their ability to faithfully represent learning dynamics remains limited. To address these challenges, the Cluster-Enhanced Behavioral Knowledge Tracing (CEBKT) model is proposed. At the core of the model, a clustering mechanism organizes exercises and interaction behaviors into a structured latent space, enabling context-aware representation learning. Built upon this structure, a behavior-aware learning gain module and a multi-factor forgetting mechanism are developed to capture fine-grained knowledge acquisition and decay processes. This enables the model to effectively characterize the interpretable evolution of students’ knowledge states over time. Experimental results on multiple public datasets show that CEBKT generally outperforms strong baselines and provides improved interpretability.