Deep knowledge tracing method based on enhancing knowledge graph embedding

Knowledge tracing is dedicated to tracing students' mastery status of knowledge points based on their responses to a sequence of questions and predicting future performance. The learning process is essentially an interaction among three types of entities: students, questions and skills. But current research methods are often limited to the analysis of direct relationships and fail to fully reveal the higher-order structural relationships among entities. Knowledge Graph is good at constructing and expressing spatial relationships among entities. However, in the application of knowledge tracing domain, some researches only focus on the relational triad of knowledge graph, but neglect the rich semantic information embedded in the attributes of each node. In this paper, we propose a deep knowledge tracing method based on enhancing knowledge graph embedding. We construct a knowledge graph based on students' historical answer records, model all three types of interactions, and then embed the three types of nodes and their attributes to capture the complex relationships among them. We perform comparison and ablation experiments on three knowledge-tracing public datasets. The model's accuracy (ACC) and area under the curve (AUC) are significantly improved, especially the best performance in EdNet dataset, which is improved by 10.2% and 9.08%, respectively, compared with DKT. This demonstrates the superiority of the method proposed in this paper for future performance prediction.