<![CDATA[Recent progress in learning analytics and educational data mining has accelerated the development of adaptive learning systems, where Cognitive Diagnostic Computerized Adaptive Testing (CD-CAT) has emerged as a significant approach. CD-CAT employs Cognitive Diagnosis Models (CDMs) to generate detailed evaluations of student competencies. Nevertheless, increasing numbers of attributes and test items introduce challenges related to the complexity and uncertainty of adaptive policies. This research investigates and compares four cognitive diagnosis models, namely GD-DINA, MIRT, MCD, and KaNCD, within an adaptive testing framework based on a Partially Observable Markov Decision Process (POMDP). The evaluation was conducted using the ASSISTments dataset with Accuracy, AUC, and expected reward as performance metrics. The findings indicate that KaNCD achieved the best overall performance, obtaining the highest diagnostic accuracy (0.7503) and AUC score (0.7410), while also maintaining stable results in POMDP-based adaptive testing (expected reward = 0.792). Although GD-DINA produced the highest expected reward (0.901), its accuracy was comparatively lower. Meanwhile, MCD demonstrated a balanced performance, with high accuracy (0.7464) and strong adaptability of policy (reward = 0.873). Overall, these results suggest that KaNCD offers the most effective balance among accuracy, interpretability, and efficiency in POMDP-based adaptive testing systems.]]>
Copyrights © 2026
