Is There a Better Way to Forget? Modelling Memory Decay in Deep Knowledge Tracing

Abstract

Knowledge acquiring and forgetting are integral components of the student learning process, both of which are critical for Deep Knowledge Tracing (DKT) tasks that leverage students’ historical learning data to forecast future performance using deep learning methods. However, forgetting mechanisms proposed in many DKT models are not grounded in adequate theories and lack clear interpretation. Building on psychological theories, this study focuses on a crucial aspect of modelling forgetting — memory decay — and performs ablation study and comparative analysis of various decay functions in DKT models to investigate the robustness of their implementation of forgetting. Our goal is to understand how forgetting is manifested in the popular DKT task, identify shortcomings, and highlight potential opportunities to guide new research in this area based on forgetting-related theories. Our results show that the wide-adopted Ebbinghaus’s forgetting curve underperforms when replaced by sigmoid and inverse decay functions in some scenarios. We also present a discussion on the factors that influence the models’ performance, including the effects of completely removing forgetting components from DKT models.

Keywords: knowledge tracing, educational data mining, learning and forgetting, memory decay, Deep Learning