Response speed enhanced fine-grained knowledge tracing: A multi-task learning perspective

Abstract

The primary objective of knowledge tracing (KT) is to trace learners’ changing knowledge states and predict their future performance by analyzing their learning trajectories. One of the fundamental assumptions underpinning KT is that estimating knowledge states is roughly equivalent to predicting future performance. However, this assumption has not been extensively explored in most studies, particularly in relation to the consistency between observable performance and latent knowledge state. To address this challenge, we propose a novel response speed enhanced fine-grained knowledge tracing (FKT) method. FKT leverages response speed through response time and integrates speed prediction as an additional task within a multi-task learning framework. Through this framework, FKT can separate representations of different knowledge state in the feature space, thereby facilitating fine-grained knowledge tracing. Moreover, we divide the task of predicting learner performance into three procedures: obtaining historical knowledge state, inferring future latent traits, and forecasting future performance. To this end, we formalize each learner’s response interaction as a time cell and develop an encoder–decoder–predictor framework for KT. To enhance the accuracy of performance prediction, we introduce a time-distance attention mechanism and knowledge proficiency component and provide two multi-task objective functions. Our experimental results on four real-world datasets demonstrate the superiority of future performance prediction and good interpretability of FKT.