A Novel and Scalable Dual Response Hypothesis Scheme Based on Transformer for Knowledge Tracing
Li Xu, Zhilong Shan · 2023
Knowledge tracing predicts learners’ performance on target questions given their past exercise performance. Although multiple relations between interactions can be exploited, their complete utilization forces the design of complex domain-specified architectures due to the heterogeneity between the target question and interaction records. A dual response hypothesis scheme is proposed to transform the KT task into a standard sequential classification task to apply the Transformer which brings about less artificial induction bias and higher capacity. The effectiveness of the approach is verified on four educational datasets. For instance, it achieves relative improvements of 2.8% on ASSISTment09, 0.9% on ASSISTment12, 10.0% on ASSISTment17, and 4.3% on EdNet. The ablation study demonstrates that the response hypothesis scheme can better utilize auxiliary relations than the domain-specific scheme.