Proceedings of the Annual Meeting of the Cognitive Science Society

Knowledge Tracing (KT) aims to predict students' future performance based on their historical interaction data. With the advancement of attention mechanisms, attention-based KT models have achieved significant improvements in predictive performance. However, most existing attention-based KT models typically employ deterministic cognitive embeddings to represent students' knowledge states. Although such representations effectively capture the stability of cognitive processes, they fail to account for the inherent volatility in students' learning behaviors, thereby limiting the capacity for modeling authentic learning dynamics. To address this limitation, we propose a novel knowledge tracing model with enhanced capability for modeling cognitive volatilities—CWKT. Specifically, we model students' interaction representations as Gaussian distributions to capture the intrinsic long-term volatility in the learning process. We further apply a frequency decomposition to the mean of the Gaussian distribution, enabling us to extract local volatility information. To model the distributional transitions of students' knowledge states during learning, we introduce a Wasserstein Self-Attention mechanism. Moreover, an attention penalty module is incorporated to mitigate the model's potential overemphasis on long-term volatility, thereby improving the overall stability of predictions. Extensive experiments conducted on four public educational datasets demonstrate that the proposed model exhibits significant advantages in capturing cognitive volatilities.