Predicting Timely Graduation for Evidence-Based Higher Education Management: An Explainable Machine Learning Approach

Higher education institutions increasingly use predictive analytics to identify students who may require academic support, yet predictive accuracy alone does not determine how model outputs should inform responsible management action. This study evaluates an explainable machine-learning framework for timely graduation decision support using routine academic records through Semester 4 and proposes a Role-Differentiated Action Translation Layer (R-ATL). Historical academic records were reconstructed into student-level trajectories, with timely graduation defined as completion by the end of the eighth semester. The analytical sample comprised 1,711 students from the 2019–2021 cohorts. Models were developed using the 2019–2020 cohorts and temporally tested on the untouched 2021 cohort. Random Forest achieved a ROC-AUC of 0.755, balanced accuracy of 0.670, and Brier score of 0.190, while XGBoost achieved a ROC-AUC of 0.761, with the small difference not statistically resolved. TreeSHAP identified a mixed performance–progression profile dominated by weighted IPS, Semester-4 IPS, early credit accumulation, and Semester-3 IPS. Informative missingness in the 2024 cohort limited complete-trajectory scoring to 438 of 604 students; these results were therefore treated as a current-cohort scoring demonstration rather than external validation. R-ATL translates explanation patterns into proposed role-specific information for students, academic advisors, program managers, and institutional leaders while retaining human review. The framework received complete content validity agreement from six experts. External validation, end-user usability and actionability testing, and intervention-outcome evaluation remain necessary for further empirical assessment and confirmation of the framework’s practical effectiveness across diverse higher education contexts and student populations