Enhancing Knowledge Tracing with Talking-Heads Attention and Item Response Theory

Abstract

Knowledge tracing (KT) is a fundamental task in student modeling within online education systems, enabling the tracking of a student’s learning progress and the prediction of future performance. Recent attention-based KT methods have leveraged self-attention mechanisms to capture complex patterns in student learning. However, conventional self-attention is constrained by independent attention heads that lack direct information exchange, limiting the model’s ability to capture dependencies across different learning interactions. To address this limitation, we introduce Talking-heads Attention into KT, enabling direct information sharing among attention heads to enhance knowledge-state modeling and improve predictive performance. In addition, to improve model interpretability, we integrate an Item Response Theory (IRT) module into the proposed model, providing a complementary interpretability framework grounded in psychometrics. Experiments on three KT benchmark datasets (ASSISTments2012, ASSISTments2017, and Junyi) demonstrate that our approach outperforms state-of-the-art KT models in prediction accuracy while preserving interpretability by integrating an IRT-informed module. Our method achieves an average 7% improvement in AUC across the three datasets.