Decoupling Ability Residuals from Affective Modeling in Cognitive Diagnosis

Summary

Cognitive diagnosis estimates students’ concept mastery from response logs, but answers are also influenced by emotion, engagement, and fatigue. The paper argues that residual errors left by conventional cognitive diagnosis models do not necessarily represent affect: item calibration bias, concept bias, student-concept deviations, and latent student-item matching can create stable cognitive patterns. It proposes an ability-residual decoupled framework that explicitly models these residuals through student, item, concept, student-concept, and low-rank student-item components. A separate affective module then modulates guess and slip effects, while a Q-matrix-constrained attention mechanism aggregates only residuals associated with concepts relevant to each item. Experiments on ASSIST2017, ASSIST2012, ASSIST2009, and Junyi use six cognitive diagnosis backbones and report improved response prediction across the reported comparisons. When affect labels are available, affect alignment generally also improves. Ablation studies, leakage probes, principal-component visualizations, long-tail analysis, and case studies indicate that the residual branch absorbs stable cognitive bias, reduces contamination of the affective branch, and improves robustness and predictive accuracy.