Efficient Attentive Knowledge Tracing for Long-Tail Distributed Records

Knowledge Tracing (KT), the most basic task for Intelligent Tutoring Systems, models individual students' knowledge state over time based on their past interactions. In large-scale KT datasets, we observe the length of student interaction records satisfy a long-tail distribution, and propose an efficient self-attentive architecture, EAKT, optimized for this situation to accelerate training. By combining linear self-attention with adaptive window size, EAKT enables the long-range modeling capability for long records and high window utilization for short records. Through experiments on two large-scale KT datasets, we demonstrate that EAKT achieves about 4 times faster with one-third memory cost than the best existing model, while still maintains competitive accuracies.