CIKT: Causality Inspired Knowledge Tracing

Shuaishuai Zu, Li Li, Songtao Cai, Jun Shen

Abstract: In intelligent systems, knowledge tracing (KT) aims to trace students’ knowledge states based on their learning interactions consisting of responses to given questions. Owing to the data sparsity problem, recent studies on KT have increasingly concentrated on incorporating as much learning-related information as possible to augment the prediction ability. However, these methods only explore such supplemental information to enhance the input representations and leave them out of consideration during the learning process. Therefore, in this paper, we propose a novel Causality Inspired Knowledge Tracing model (CIKT), which discovers the causal relationship between knowledge concepts from learning interactions and integrates such intrinsic relations for the prediction task. Specifically, two types of cause-effect graphs are first constructed by using the Granger causality test. In addition, a directed information transmission mechanism is proposed to propagate the causal relationships into each knowledge concept with triplet representations: two types of cause representations and one type of effect representation. Then, a mask attention mechanism is proposed to differentiate effects from different types of causes that relate the students’ performance to their past interactions. More importantly, we reveal that incorporating the causal relationships into the masked attention mechanism can reduce the inference uncertainty, based on which we give theoretical analysis of the effectiveness. Experiments on six real-world datasets show that our model’s advanced predictive performance over other baselines.