From Prediction to Targeting: Comparative ML Models and Threshold-Based Re-Enrollment Segmentation in Higher Education Marketing

Student re-enrollment is a critical strategic concern for higher education institutions, directly impacting financial stability, capacity planning, and the effectiveness of retention programs. This study develops a decision-support approach to predict non–re-enrollment using institutional records from a private university in Central Java. The dataset includes 2,673 student records across three academic years, which were expanded into 1,099 engineered features through a unified preprocessing pipeline involving missing-value imputation, scaling, and one-hot encoding. Model development utilized a fixed train–test split with 5-fold cross-validation. Results demonstrate that tuned tree ensembles significantly outperform the linear baseline. While Logistic Regression yielded limited discrimination (test ROC-AUC = 0.5602, F1 = 0.7010), tuned Random Forest improved classification quality (test ROC-AUC = 0.7571, F1 = 0.8052). Tuned XGBoost achieved the strongest ranking performance (test ROC-AUC = 0.7606) and was selected for deployment due to its superior risk-ordering capability. SHAP-based interpretation identifies parental income as the dominant driver of non–re-enrollment risk, followed by program-choice indicators and demographic variables. Finally, threshold analysis supports risk-tier segmentation, translating predicted probabilities into practical outreach policies aligned with institutional capacity constraints—addressing two underexplored gaps in applied re-enrollment prediction: rigorous cross-validated ensemble modeling and the integration of predictive scores into actionable marketing segmentation, and highlighting that ranking quality—not classification alone—is essential for operational targeting.