Explainable AI in educational data mining: LightGBM and SHAP for enhancing learner performance prediction

Abstract

Accurate prediction of learner performance is essential for adaptive educational systems and personalized learning. However, the integration of advanced gradient-boosting models with established educational frameworks remains underexplored, particularly when predictive performance and interpretability are considered jointly. This study evaluates Light Gradient Boosting Machine (LightGBM) within three established learner-modeling frameworks—Item Response Theory (IRT), Performance Factor Analysis (PFA), and Dynamic Aspects of Student Skill with History (DAS3H)—across seven real-world educational datasets. LightGBM is compared with traditional logistic regression using the Area Under the Receiver Operating Characteristic Curve (AUC) and accuracy as evaluation metrics. To further assess model interpretability, SHapley Additive exPlanations (SHAP) are employed to identify the contributions of user-, item-, and skill-level features to learner performance predictions. Results show that LightGBM generally improves predictive performance compared with logistic regression, particularly in capturing nonlinear relationships, temporal patterns, and skill-dependent learner behaviors. The SHAP analysis further provides insights into the relative importance of different feature types, enhancing the interpretability of the predictive models. These findings demonstrate that integrating LightGBM with established learner-modeling frameworks can provide a more accurate and interpretable approach to learner performance prediction. The novelty of this study lies in combining LightGBM with IRT, PFA, and DAS3H, together with SHAP-based explainability, across multiple educational datasets, offering a comprehensive framework for accurate and interpretable learner modeling in adaptive and personalized learning environments.