Abstract: Knowledge tracing is key to adaptive learning,aiming to assess students’ knowledge states and predict their future performance.Currently,data sparsity limits existing knowledge tracing models in both question embedding learning and student knowledge state modeling.To address this,some studies have introduced contrastive learning.However,existing contrastive learning methods rely on random perturbations of graph structures(for question embedding) or modifications of learning interaction sequences(for knowledge state modeling) to generate contrastive views,introducing noise and erroneous self-supervised signals.This results in question embeddings that are poorly suited for downstream tasks in learning systems.To overcome these li-mitations,this study proposes an innovative Dual-level Contrastive Learning Framework(DCLF) to simultaneously enhance question embedding learning and student knowledge state modeling in knowledge tracing.DCLF adopts a more effective contrastive paradigm that avoids altering the original data information.Instead,it generates contrastive views through relational transformations of the original data or by leveraging outputs from different neural networks on the same data.Specifically,for embedding learning,the proposed method obtains contrastive views through relational transformations of the data.For student knowledge state modeling,it encodes learning interactions using different neural networks to obtain knowledge states under various encoders.This method extracts rich self-supervised signals from multiple contrastive views,preserving the intrinsic semantic information of the data and effectively avoiding noise introduction.Experiments conducted on three commonly used datasets demonstrate that DCLF outperforms selected existing knowledge tracing models in terms of performance.

Key words: Knowledge tracing, Adaptive learning, Contrastive learning, Contrastive paradigm, Self supervised learning