Behavior-Augmented Knowledge Tracing with Monotonic Attention and Cross-Feature Fusion

Knowledge tracing is a critical research area in online education, which aims to model students’ mastery of knowledge by analyzing their historical response data and to predict their future performance accordingly. Although significant progress has been made in this domain, most existing models primarily focus on using students’ question-answering sequences as input, while largely neglecting other valuable information available on online learning platforms. To address this limitation, this paper proposes a novel Behavior-Augmented Knowledge Tracing model BAKT, which leverages a monotonic attention mechanism and cross-feature fusion to incorporate enriched behavioral signals for more accurate student performance modeling. BAKT integrates multiple types of student behavioral data through cross-feature fusion, including the number of hints requested, the number of attempts made, and the time spent on answering each question. Finally, a recurrent neural network (RNN) is employed to incorporate these behavioral features, enhancing the representation of students’ knowledge states. Experimental results on four benchmark datasets demonstrate that BAKT achieves superior performance, validating its robustness and effectiveness in the knowledge tracing task.