A Deep Memory-Aware Attentive Model for Knowledge Tracing

With the rapid development of Internet technology, online learning has become a major learning method for students. As a key component of online learning, the knowledge tracing mainly predicts students' performance in future questions by tracing the changes of students' knowledge mastery during the learning process. According to our research, most current deep knowledge tracing models take more into account factors such as students' response time, number of questions, and difficulty of questions, ignoring some aspects of psychological research on students' forgetting behaviors. Secondly, although the current knowledge tracing models based on attention mechanism are better than the traditional methods in performance prediction, they cannot capture the complex relationship between questions and responses over time because the attention layer is too shallow. In the paper, we propose a novel deep memory-aware attentive knowledge tracing model named DMAKT based on Transformer architecture. DMAKT has an encoder-decoder structure where the question embedding sequences enter the self-attention layer of the encoder, and the response embedding sequences as well as the question embedding sequences enter the self-attention layer of the decoder. DMAKT can capture the complex relationship between questions and responses through deep self-attention. Furthermore, DMAKT proposes a novel encoder-decoder monotonic attention mechanism for emphasizing performance information on recent related and other questions by considering the time interval factor of students' forgetting behavior. Finally, our extensive experiments on three real datasets show that DMAKT model improves the AUC value by an average of 4.9% compared with existing deep knowledge tracing models.