<![CDATA[Student graduation rate is a critical performance indicator for higher education institutions, particularly in accreditation assessment. Early prediction of on-time graduation supports academic planning and quality assurance. Although prior studies report high predictive accuracy using conventional cross-validation, limited attention has been given to robustness under cohort-based distribution shifts. This study evaluates the robustness of three gradient boosting models—Histogram-Based Gradient Boosting (HGB), Extreme Gradient Boosting (XGBoost), and Light Gradient Boosting Machine (LightGBM)—for predicting on-time graduation using structured academic trajectory data from 370 labeled instances across three cohorts. Two validation strategies were employed: a stratified 80:20 split for aggregated evaluation and Leave-One-Group-Out (LOGO) validation to simulate cohort-based distribution shifts. Under stratified evaluation, all models achieved macro F1-scores above 0.74, with HGB obtaining the highest score (0.7568). However, LOGO evaluation revealed substantial performance degradation, with mean F1-scores below 0.51 and increased variability across cohorts. XGBoost demonstrated comparatively better stability under distribution shifts. These findings indicate that high predictive accuracy under random splits does not guarantee cross-cohort robustness, highlighting the importance of distribution-aware validation for reliable deployment in educational data mining.]]>
