Explainable AI Frameworks for Educational Recommendation Systems

Abstract

Educational recommendation systems -- which suggest learning content, pathways, and resources to learners -- increasingly rely on opaque machine learning models whose recommendations are difficult for learners, instructors, and administrators to understand or trust. Explainability is particularly critical in educational contexts where recommendations affect learning trajectories, where learner agency and metacognitive development depend on understanding why content is recommended, and where regulatory frameworks (GDPR Article 22, EU AI Act) mandate explainability for automated decision-making affecting individuals. This paper proposes the Explainable Educational Recommendation Framework (XEDF), a systematic evaluation of six explainable AI (XAI) approaches -- SHAP, LIME, attention visualisation, counterfactual explanations, pedagogical rule extraction, and natural language explanation generation -- applied to four educational recommender architectures across three stakeholder perspectives: learner, instructor, and system administrator. XEDF evaluates explanation quality on 1,800 recommendation instances using the Explanation Utility Score (EUS) integrating fidelity, comprehensibility, and actionability. Key results: natural language generation achieves the highest learner EUS (0.912); pedagogical rule extraction achieves the highest instructor EUS (0.884); SHAP achieves the highest administrator fidelity (0.942). The framework provides XAI method selection guidance for educational recommender system designers.