Graph Contrast Learning Fused Question Features Based Knowledge Tracing

The exponential growth of online education has amplified the need for personalized learning, with Intelligent Tutoring Systems (ITS) providing rich behavioral interaction data for adaptive instruction. However, accurately assessing learners' knowledge states remains a significant challenge. Knowledge Tracing (KT), which predicts student performance based on historical interactions, has emerged as a promising solution. The complexity of KT lies in the multidimensional and dynamic nature of knowledge acquisition. In Graph-based Knowledge Tracing (GKT), current models face limitations in node importance representation, question embedding sparsity, and underutilization of learning features. To this end, this paper proposes the Graph Contrast Learning Fused Question Features Based Knowledge Tracing (GCLFKT) model. Firstly, GCLFKT employs Graph Attention Networks (GAT) to generate aggregated question embeddings, mitigating data sparsity problems associated with question-level embeddings. Furthermore, it incorporates Graph Contrastive Learning (GCL) to enhance embedding representations and integrates rich question features through a feature fusion layer. Experiments on the ASSISTments datasets demonstrate that GCLFKT outperforms several baseline models by over 4% in AUC metrics, with superior predictive performance validated through visualization analysis.