Multi-Scaled Attentive Knowledge Tracing

Learning is a dynamic, complex, and time-series process. Knowledge tracing (KT) aims to simulate learners' learning process by using learners' behavioral performance in past learning activities. In recent years, self-attentive mechanisms have been widely used in KT model. The literature shows that attention-based KT models generally outperform traditional deep knowledge tracing models. In order to simulate the learning process of learners more effectively we propose a new multi-scale attentive knowledge tracing model for KT. Specifically, the model uses multi-scale multi-head attention to capture learner features at different time scales and use them to model learners' learning behaviors. We also use relative position encoding to maintain the consistency of location information across multiple scales of attention. Experiments on real datasets show that our model outperforms state-of-the-art KT methods.