Graph-based Dynamic Interactive Knowledge Tracing

Abstract

Knowledge Tracing (KT) is a research field that aims to trace the students' knowledge states based on their historical learning. Much research has explored the value of relations among concepts and proposed to introduce knowledge structure into KT tasks. However, these studies suffer from two major shortcomings: 1) they only consider the dynamic state of students or the stationary property of concepts, but ignore the dynamic state of concepts and questions. 2) they do not make full use of the rich structural information of the relations between questions and concepts. In this paper, we propose a Graph-based Dynamic Interactive Knowledge Tracing (DGKT) to address these limitations. Specifically, we design two embeddings for each student, question, and concept: static embedding and dynamic embedding, which respectively represent stable attributes and time-varying properties. Additionally, DGKT utilizes graph self-supervised learning to enrich the stationary embeddings of questions and concepts. Finally, extensive experiments on three real-world datasets demonstrate that the proposed DGKT outperforms other baseline models.