Séminaire « Evaluation Framework for Explainable AI in Learning Analytics » – 9 février 2026

Intervenante : Linda Benachenhou

Résumé : The growing use of predictive models of learning performance in Learning Analytics (LA) raises critical concerns about the intelligibility, interpretability and trustworthiness of algorithmic decisions. Although Explainable AI (XAI) methods have been introduced to make these models more transparent, there is still no standardized approach to evaluate the quality of the explanations they provide. In this work, we propose a comprehensive evaluation framework, applied on the Open University Learning Analytics Dataset (OULAD), comparing several widely used XAI methods, including SHAP, LIME, and counterfactual explanation. Our results reveal variability in explanation quality across methods, demonstrating the need for multi-faceted evaluation approaches.