Knowledge structure enhanced graph representation learning model for attentive knowledge tracing

Knowledge tracing (KT) is a fundamental personalized-tutoring technique for learners in online learning systems. Recent KT methods employ flexible deep neural network-based models that excel at this task. However, the adequacy of KT is still challenged by the sparseness of the learners' exercise data. To alleviate the sparseness problem, most of the exiting KT studies are performed at the skill-level rather than the question-level, as questions are often numerous and associated with much fewer skills. However, at the skill level, KT neglects the distinctive information related to the questions themselves and their relations. In this case, the models can imprecisely infer the learners' knowledge states and might fail to capture the long-term dependencies in the exercising sequences. In the knowledge domain, skills are naturally linked as a graph (with the edges being the prerequisite relations between pedagogical concepts). We refer to such a graph as a knowledge structure (KS). Incorporating a KS into the KT procedure can potentially resolve both the sparseness and information loss, but this avenue has been underexplored because obtaining the complete KS of a domain is challenging and labor-intensive. In this paper, we propose a novel KS-enhanced graph representation learning model for KT with an attention mechanism (KSGKT). We first explore eight methods that automatically infer the domain KS from learner response data and integrate it into the KT procedure. Leveraging a graph representation learning model, we then obtain the question and skill embeddings from the KS-enhanced graph. To incorporate more distinctive information on the questions, we extract the cognitive question difficulty from the learning history of each learner. We then propose a convolutional representation method that fuses these disctinctive features, thus obtaining a comprehensive representation of each question. These representations are input to the proposed KT model, and the long-term dependencies are handled by the attention mechanism. The model finally predicts the learner's performance on new problems. Extensive experiments conducted from six perspectives on three real-world data sets demonstrated the superiority and interpretability of our model for learner-performance modeling. Based on the KT results, we also suggest three potential applications of our model.