<p>Knowledge tracing seeks to model and predict learners’ evolving knowledge states based on their historical learning interactions. Traditional models primarily rely on question-answering behaviour and often overlook the role of affective states in learning. This study proposes a multimodal cross-attention knowledge tracing model (MCA-KT) that integrates facial expression features with behavioural data collected during students’ answering processes. A cognitively guided cross-attention mechanism is introduced, in which behavioural representations actively select and weight facial expression features, explicitly modelling the directional modulation of affect by cognition during learning. This design is grounded in affective-cognitive interaction theory, which emphasises the bidirectional influence between emotional states and cognitive processes. A masking mechanism is further incorporated to handle missing expression data and enhance model robustness. Seven baseline models from statistical, deep learning, and Transformer-enhanced categories are selected for comparison. Experiments on three extended datasets show that MCA-KT consistently achieves higher AUC than all baselines, and ablation studies verify the effectiveness of the cross-attention mechanism and the masking strategy. Overall, the results demonstrate that explicitly modelling affect–cognition interactions via cross-attention leads to more accurate and robust knowledge tracing. The proposed MCA-KT provides an effective multimodal modelling paradigm for incorporating affective signals into sequential student modelling.</p>