Unleashing the potential of un-interacted exercises for boosting cognitive diagnosis

Cognitive diagnosis serves as a cornerstone of intelligent education, providing detailed insights into students’ knowledge acquisition. However, foundational methods in cognitive diagnosis heavily rely on interaction logs, which are frequently sparse and thus constrain the accuracy and effectiveness of diagnosis. Although some approaches attempt to alleviate the sparsity problem by accounting for information incompleteness or enhancing knowledge graph modeling, they still primarily center around interaction-based information, largely overlooking the rich diagnostic potential embedded in un-interacted exercises. We propose the Un-Interacted-Enhanced Cognitive Diagnosis (UECD) framework to address these challenges by incorporating information from un-interacted exercises. Specifically, the framework employs a two-stage dynamic sampling strategy that systematically integrates student characteristics, exercise attributes, and knowledge concepts under specific constraints to select valuable un-interacted exercises. An attention mechanism is introduced to capture relationships between interacted and un-interacted exercises, enabling adaptive information fusion. Additionally, a pseudo-label generation module, guided by the consistency assumption and a confidence-based filtering mechanism, ensures the reliability of generated labels. The framework is model-agnostic and seamlessly compatible with most cognitive diagnosis models. Experimental evaluations on four real-world datasets demonstrate that UECD significantly improves diagnostic accuracy, mitigates data sparsity, and enhances the adaptability of cognitive diagnosis tasks.