Effective student performance prediction enables timely support in Educational Data Mining (EDM). However, most predictive models lack interpretability and do not provide actionable feedback for learners. To address prediction accuracy and explanation, this study integrates an optimized XGBoost classifier with SHapley Additive exPlanations (SHAP). Using the Open University Learning Analytics Dataset (OULAD), our model employs SMOTE-based class balancing and randomized hyperparameter tuning, achieving better performance than existing research. Moreover, the model supports early predictions using demographic, assessment, and behavioral data, making it suitable for timely interventions. To address the challenge that predictive models typically function as black boxes without providing concrete guidance for instructors and students, SHAP explanations generate individualized reports that highlight the most influential features for each student’s predicted outcome. These explanations are mapped to targeted, prioritized feedback aligned with students’ learning needs. Case studies demonstrate how the system offers concrete and interpretable guidance for both at-risk and high-performing students. This study addresses a key EDM gap by moving beyond prediction to support personalized intervention. By linking SHAP-based explanations to personalized feedback, our method provides a scalable and interpretable framework for targeted support in education.