An Explainable Educational Data Mining Framework for Misconception Detection and Differentiated Instruction

Abstract

Artificial Intelligence (AI) and Generative Artificial Intelligence (GenAI) increasingly support personalised learning, intelligent assessment, and data-informed educational decision-making. However, persistent misconceptions among gifted students may remain undetected because high achievement can conceal conceptually coherent misunderstandings. This study develops and empirically evaluates a teacher-governed, closed-loop framework integrating Educational Data Mining (EDM), Learning Analytics (LA), ensemble machine learning, Explainable Artificial Intelligence (XAI), GenAI, and differentiated instruction. Educational data from 255 gifted secondary-school learners were transformed into learner profiles and analysed using Random Forest, XGBoost, Boost, Support Vector Machine, Artificial Neural Network, and Logistic Regression, which were then combined via validation-derived weighted probabilistic ensembling. The validation-weight ensemble achieved 96.88% accuracy, 92.00% precision, 100% recall, an F1-score of 95.83%, ROC-AUC of 0.9979, and PR-AUC of 0.9963, outperforming the individual classifiers in overall classification performance. SHAP provided learner-specific explanations, with average response time emerging most frequently as the leading explanatory feature. These explanations informed GenAI-supported personalised interventions, which teachers validated before differentiated delivery. Post-intervention evidence was subsequently incorporated into learner reassessment and profile updating. The principal contribution lies in the operational integration of learner profiling, weight prediction, explainability, evidence-conditioned instructional generation, professional validation, and reassessment within a continuous educational cycle. The findings provide initial empirical evidence for the framework’s computational feasibility and interpretability while supporting further external and longitudinal validation across educational contexts and learner populations.