Graph Knowledge Structure for Attentional Knowledge Tracing With Self-Supervised Learning

As intelligent education advances and online learning becomes more prevalent, Knowledge Tracing (KT) has become increasingly important. KT assesses students' learning progress by analysing their historical performance in related exercises. Despite significant advances in the field, there are still shortcomings in two aspects: first, a lack of effective integration between exercises and knowledge points; second, an overemphasis on nodal information, neglecting deep semantic relationships. To address these, we propose a self-supervised learning approach that uses an enhanced heterogeneous graph attention network to represent and analyse complex relationships between exercises and knowledge points. We introduce an innovative surrogate view generation method to optimise the integration of local structural information and global semantics within the graph, addressing relational inductive bias. In addition, we incorporate the improved representation algorithm into the loss function to handle data sparsity, thereby improving prediction accuracy. Experiments on three real-world datasets show that our model outperforms baseline models.