From Prediction to Support: Explainable Machine Learning and Rule-Based Interventions for Identifying Students at Academic Risk

Early identification of students at risk of academic failure may enable institutions to offer timely and proportionate support. This study developed and compared three supervised classification models, XGBoost, Random Forest and Decision Tree, and linked their predictions to a prototype rule-based intervention dashboard. A public synthetic student-performance dataset containing 6,607 records was preprocessed by imputing missing values and encoding categorical variables. Academic risk was defined by an examination-score threshold, with scores below 65 classified as high risk and scores of 65 or above as low risk. Data were divided into 80% for training and 20% for testing. Models were assessed using accuracy, precision, recall, F1-score, confusion matrices and area under the receiver operating characteristic curve. SHAP values were used to examine the direction and relative importance of predictors. XGBoost achieved the strongest test performance, with 0.92 accuracy, 0.92 weighted precision, 0.92 weighted recall, 0.92 weighted F1-score and an AUC-ROC of 0.976. Random Forest achieved 0.91 accuracy and an AUC-ROC of 0.965, while Decision Tree achieved 0.83 accuracy and an AUC-ROC of 0.809. For the high-risk class, XGBoost produced a precision of 0.89 and recall of 0.86. Across the models, hours studied, attendance, and previous scores were the most influential features. These informed a Power BI prototype that paired risk classifications with suggested support. Findings show technical feasibility of combining explainable prediction with recommendations. However, because the dataset is synthetic, results do not establish effectiveness in a university population. External validation, fairness assessment, and human oversight are required before operational use.