What is wrong with deep knowledge tracing? Attention-based knowledge tracing

Scientifically and effectively tracking student knowledge states is a significant and fundamental task in personalized education. Many neural network-based models, e.g., deep knowledge tracing (DKT), have achieved remarkable results on knowledge tracing. DKT does not require handcrafted knowledge and can capture more complex representations of student knowledge. However, a severe problem of DKT is that the output fluctuates wildly. In this paper, we utilize a finite state automaton (FSA), a mathematical computation model, to interpret the waviness of DKT because an FSA has observable state evolution in response to external input. With the support of an FSA, we discover that DKT cannot handle long sequential inputs, which leads to unstable predictions. Accordingly, we introduce two novel attention-based models that solve the above problems by directly capturing the relationships among each item of the input sequence. Extensive experimentation on five well-known datasets shows that our two proposed models achieve state-of-the-art performance compared to existing knowledge tracing approaches.