Integrating learning factors and Bayesian network for interpretable knowledge tracing

Abstract

Knowledge tracing aims to model learners’ knowledge mastery based on their historical interaction records and predict their future performance. Due to its great potential in enabling personalized learning in intelligent tutoring systems, it has received extensive attention. However, most deep learning-based knowledge tracing methods have significant predictive performance. It is difficult to extract meaningful interpretations from the thousands of parameters in neural networks. The interpretability of knowledge tracing refers to the ability of learners to easily understand the predicted results.To address this problem, based on learning factors that influence the learner’s exercise performance, this paper proposes a novel knowledge tracing model which is named Integrating Learning factors and Bayesian network for interpretable Knowledge Tracing (LBKT). Firstly, meaningful learning factors, including knowledge mastery, learning ability, and exercise difficulty, are calculated from learners’ historical interaction records using deep learning and statistical methods. Then, Bayesian network is constructed to capture the causal relation between the three learning factors and exercise response. Finally, the Bayesian network is generated through structure and parameter learning to obtain interpretable prediction of future exercise performance. The proposed model named LBKT is evaluated on three public real-world educational datasets. The experiment results demonstrate that our approach achieves better predictive performance compared to baseline knowledge tracing methods, while also exhibiting significant superiority in model interpretability.