Predicting On-Time Graduation in Indonesian Private Higher Education: A Machine Learning Comparison for Institutional Decision Support

Timely graduation shapes accreditation and resource decisions in Indonesia's pri-vate higher education, so how predictive models are built and used is a social as much as a technical question. Prior machine learning comparisons in this setting rely on overall accuracy, run under differing conditions, and rarely test how well models detect late-graduating students or inform institutional action. This study compares Random Forest (RF) and K-Nearest Neighbor (KNN) for on-time gradua-tion prediction on 656 student records (2018–2022) under identical conditions, using accuracy, precision, recall, F1-score, 10-fold cross-validation, feature im-portance (MDI), and the Wilcoxon Signed-Rank Test with a rank-biserial effect size. RF reached higher accuracy (80.30% versus 70.45%), but both models detected the minority late-graduating class poorly (recall 0.40 and 0.26), and the difference was not statistically significant (p = 0.3477; r ≈ 0.49). Academic attributes accounted for 72.38% of feature importance. These results position such models as accountable decision support rather than a validated early-warning intervention, which would require evaluation in real academic governance.