Dual-State Knowledge Tracing Model with Mutual Information Maximization

Knowledge tracing aims to trace students' knowledge states and predict their future performance based on their historical learning processes. Most existing methods of characterizing a student's state are not effective enough, using only global representation or knowledge concept level representation. Such representation methods cannot consider the characteristics of knowledge concepts and the relations between concepts at the same time. In this paper, we propose a Dual-State Knowledge Tracing (DSKT) Model with Mutual Information Maximization. DSKT uses dynamic routing to extract knowledge commonalities from original knowledge concepts, updates the knowledge state at the concept and commonality levels, and predicts future performance by fusing two states. In addition, to incorporate the relationship between exercises and knowledge concepts, we use the principle of mutual information maximization to learn their representations. Extensive experimental results show the effectiveness of our model.