CADKT: A Cognitive-Affective Dual-Graph Contrastive Framework for Dynamic Affect-Aware Knowledge Tracing

Knowledge tracing (KT) predicts students’ future performance by modeling their knowledge mastery from historical learning interactions. Although recent deep learning-based KT models have achieved promising results, most of them mainly rely on response sequences or cognitive structures, leaving affect-related learning states insufficiently explored. In real learning scenarios, student performance is shaped not only by cognitive mastery but also by affect-related factors that change during the learning process. To address this issue, this paper proposes a Cognitive-Affective Dual-Graph Knowledge Tracing framework, termed CADKT, which jointly models cognitive structure and behavior-derived affect-related dynamics. Specifically, CADKT first extracts affect-related behavioral factors from learning logs, including correctness discrepancy, time efficiency discrepancy, and learning persistence deviation. These factors are then used to infer interval-level behavior-derived affect-related profiles without relying on explicit emotion annotations, and the inferred profiles are organized by a Temporal Affective Graph to model their transitions along the learning sequence. Meanwhile, a Heterogeneous Knowledge Graph is constructed to represent student–problem interactions and problem–concept associations. To integrate the two views, CADKT introduces cross-view contrastive regularization and a Projected Hybrid Softmax fusion mechanism for next-response prediction. Experiments on four public educational datasets show that CADKT achieves stronger overall performance than representative KT baselines on most evaluation metrics. Further ablation studies, profile analysis, and case visualization demonstrate that CADKT improves prediction performance while providing interpretable evidence for the interaction between knowledge mastery and behavior-derived affect-related dynamics.