GCN and GAT-based interpretable knowledge tracing model

Abstract

Knowledge tracing (KT) aims to predict students’ future performance by assessing their level of knowledge mastery from past problem-solving records. However, many existing methods fail to take full advantage of the potential relationship between questions and skills, or fail to effectively utilize students’ historical learning data, which makes it difficult to accurately capture individualized mastery for each question. In addition, redundancy in long sequential information often leads to model overfitting, and existing deep knowledge tracing models have significant limitations in terms of the interpretability of their predictions. To address these issues, we propose GCAKT, an interpretable KT model focuses on student problem mastery. GCAKT generates personalized problem representations by modeling students’ historical learning information at a fine-grained level, and learns these representations jointly through a problem-skill embedding module and a personalized problem mastery module. To cope with the redundancy generated by long sequences, GCAKT employs an attention-based knowledge evolution module that constructs a final hidden knowledge state by analyzing the attention relationship between the student’s hidden knowledge state and the problem at each point in time. Meanwhile, GCAKT utilizes the attention weights to construct interpretable paths, aiding to provide interpretable prediction results. Experimental results on three publicly available real-world educational datasets show that GCAKT outperforms traditional methods in terms of both prediction accuracy and interpretability.