Abstract
Cognitive diagnosis, pivotal for AI-enhanced learning, faces the challenge of fusing robust modeling of complex student-exercise-knowledge relationships with psychometric theoretical consistency: (1) performance degradation from over-smoothing induced by polarized knowledge associations in graph attention networks, and (2) limited interpretability due to the opaque nature of deep learning architectures. We present GEAR-CD, a GCN-enhanced graph attention framework with adaptive relation pruning that addresses these limitations through three key innovations: First, hierarchical graph attention unifies interaction modeling across heterogeneous relationships (knowledge concepts, exercises, and learners). Second, GCN-based convolutional kernels with automated edge pruning mitigate attention-driven over-smoothing. Third, theoretically-grounded design ensures alignment with Item Response Theory (IRT) principles. Comprehensive evaluations on Junyi, Assistments and EdNet datasets demonstrate GEAR-CD’s superiority, achieving 6.85% and 2.05% relative improvements over baseline and state-of-the-art models respectively in prediction accuracy (ACC=81.67%), while maintaining exceptional stability ([Formula: see text]0.0001). Visualization analyses confirm both diagnostic robustness (via radar plots) and theoretical validity (t-SNE manifolds linearity [Formula: see text]=0.93), establishing GEAR-CD as a credible solution for operational learning systems. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1038/s41598-025-29367-7.