Sequential contrastive learning for progressive knowledge tracing

Abstract

In recent years, knowledge tracing has received significant attention in personalized education. It dynamically assesses users’ knowledge states based on their historical response sequence. User response sequences are central to knowledge tracing. While most studies focus on modeling short-term and long-term dependencies, few consider the order in which interactions occur. A recent study argues that the interaction order has little impact on users’ knowledge states (Lee et al., The Web Conference, 2022), which contradicts both our intuition and constructivist learning theory. To address this contradiction, we propose a Sequential Contrastive Learning algorithm for Progressive Knowledge Tracing, termed SPKT, to test the effectiveness of order information within the response sequences for assessing users’ knowledge states. SPKT embeds order information into the response sequence representation through a carefully designed contrastive learning module, and captures users’ monotonic memory decay patterns using a carefully designed non-symmetrical augmented view construction method. The enhanced sequence representation is subsequently utilized to decode user behavior with a progressive learning process module. Extensive experiments demonstrate that, on average, SPKT outperforms 10 baselines by up to 14 % in AUC and 8 % in ACC across 6 real-world datasets. Furthermore, the results highlight that the order information in response sequences significantly improves algorithmic performance-sometimes even more than the correctness of the responses themselves. Moreover, SPKT more accurately evaluates users with better academic performance and shorter learning sequences. For the same user, longer response sequences are more helpful in assessing a user’s knowledge state.