Modeling Student Knowledge Development Through Hierarchical Attention and Adaptive Learning Signals

Abstract

Understanding how students acquire and retain knowledge is essential for building effective personalized learning systems. Traditional knowledge tracing models primarily infer student proficiency from sequences of correct and incorrect responses, often overlooking differences in learning speed, concept dependencies, and engagement behavior. This study proposes a hierarchical attention architecture that jointly models question-level interactions, concept-level dependencies, and individual learning trajectories. The framework incorporates response accuracy, time intervals, repeated attempts, and exercise difficulty to generate personalized estimates of future learning performance. An experimental dataset is constructed with 7.4 million learning interactions from 52,600 students across 320 instructional modules. Evaluation against recurrent knowledge tracing and transformer-based baselines shows that the proposed approach achieves an AUC of 0.861 for predicting subsequent response correctness, improving upon the strongest baseline of 0.824. The model also reduces prediction error for students with limited interaction histories by 18.6%. In a simulated adaptive exercise selection experiment, the proposed strategy improves concept mastery scores by an average of 9.3 percentage points after 20 learning sessions. These results indicate that hierarchical behavioral modeling could provide a useful foundation for personalized digital education.