Exercise Feature-Enhanced Knowledge Tracing

The goal of Knowledge Tracing (KT) is to accurately assess a learner's knowledge state by analyzing and modeling learning activity data. In recent years, KT models based on deep neural networks have exhibited outstanding performance. However, they still face two major limitations: (1) Existing exercise feature construction methods predominantly rely on expertlabeled information, neglecting to uncover potential associations between knowledge concepts from the learner's perspective. (2) Given the vast yet sparse learner-exercise interaction data, it is challenging for existing models to balance the commonalities and differences among exercises. This paper proposes a knowledge tracing model named EFEKT to construct high-quality exercise feature representations. EFEKT explores potential associations between knowledge concepts from the learner's perspective and incorporates a graph attention network to extract knowledge structure information. Subsequently, it employs vector quantization to reconstruct exercise features and designs a difficulty recovery task to further optimize exercise feature representations. Designed as a modular component, the EFEKT model allows for integration with existing deep KT models. Extensive experiments conducted on two commonly used datasets confirm its superiority.