Boosting Knowledge Tracing With Structure-Preserving Adaptive Contrastive Learning

Knowledge tracing (KT) is a vital task in intelligent tutoring systems, aiming to assess learners’ knowledge states by analyzing their learning interactions. However, current KT research faces two significant challenges: (1) The exercise representations derived through Contrastive Learning (CL) often suffer from misguidance due to the use of random perturbations, and fail to capture long-distance dependencies that are closely related but distant within the knowledge structure. (2) The impact of learning efficiency on KT has not been adequately addressed. To address these issues, this paper proposed a novel KT model, named Cross-view contRastive leArning facilitatiNg Efficiencyaware knowledge tracing (CRANE)1. This model is designed to explore high-quality exercise representations and the impact of dynamically changing learning efficiency. Specifically, (1) to tackle the first challenge, a contrastive paradigm is introduced based on the generated relationship graphs to enhance the effectiveness of self-supervised signals, while an adaptive sampling method uncovers long-distance dependencies among exercises. (2) To address the second challenge, an efficiency gate is proposed to capture periodic learning gains, thereby quantifying learning efficiency and its impact on the learning process. Experiments on three common datasets demonstrate the superiority of our model and confirm the effectiveness of CRANE through visualization of exercise representations.