QSDKT: Graph-Based Dynamic Knowledge Tracing through Question-Skill Similarity

Knowledge tracing aims to trace learners’ evolving knowledge states by predicting their future performance. In the teaching and learning processes, question difficulty is usually related to knowledge skills, and the effects of question difficulty and skill difficulty on students’ knowledge states are dynamic. In some recent studies, the effect of difficulty in questions has been considered, but dynamically assessing the difficulty relationship between questions and skills has not been explored and utilized effectively. In this paper, we propose a graph-based dynamic knowledge tracing model through question-skill similarity, named QSDKT. The model is used to address the challenge of effectively capturing learners’ dynamic knowledge states. First, we construct a relation graph and define difficulty levels based on the similarity between questions and skills. Second, we employ graph convolutional networks to capture the complex relationships between questions and skills. Finally, we design an adaptive sequential learning network to adjust question difficulty and track learners’ knowledge states dynamically. Experiments on three real-world datasets validate the effectiveness of QSDKT. The results show that QSDKT achieves higher accuracy in predicting the knowledge states of students, and the adaptive network’s dynamic adjustment of the difficulty of the questions further enhances the personalized learning of the model.