Explainable At-Risk Student Prediction with SHAP-Guided Risk Segmentation and Rule-Based Intervention Recommendations

Early identification of students at risk of failure or withdrawal can support timely academic assistance, but many predictive systems provide risk labels without transparent reasons or guidance on suitable support actions. An explainable at-risk student prediction workflow combined Random Forest and XGBoost classification, SHAP-based explanation, k-means risk segmentation, and rule-based intervention recommendations. The primary evaluation used Open University Learning Analytics Dataset records observed at day 60, with the 2014J presentation retained as a locked temporal holdout; Kalboard 360 provided a secondary stratified-cross-validation robustness check. XGBoost was selected during development and achieved accuracy of 0.679, precision of 0.554, recall of 0.802, F1-score of 0.656, ROC-AUC of 0.798, and PR-AUC of 0.747 on the OULAD temporal holdout. The recall-oriented threshold identified 2,695 of 3,359 at-risk students, while producing 2,167 false-positive alerts requiring staff review. SHAP identified assessment completion ratio, mean assessment score, available weighted score, and last active day as the leading global predictors. Clustering predicted at-risk students into three groups produced a silhouette score of 0.409 and stability ARI of 1.000; the resulting profiles were mapped to tutoring, study-skills support, engagement monitoring, and academic-adviser review. The findings show that ensemble prediction can be connected to transparent risk profiles and practical support recommendations. However, the workflow remains a decision-support prototype requiring local validation and human oversight.