Delayed student graduation is a critical issue in higher education because it affects academic planning, student support, and institutional performance evaluation. This study develops a leakage controlled machine learning framework for early identification of students at risk of delayed graduation, using academic records available through the sixth semester. A dataset of 564 students was used, with graduation status defined as on-time for students graduating in the eighth semester or earlier and delayed for those graduating after the eighth semester. To prevent temporal data leakage, post-outcome variables were excluded from the predictor set. Five supervised learning algorithms were evaluated: Decision Tree, Support Vector Machine, Random Forest, Naïve Bayes, and K-Nearest Neighbor. Preprocessing was performed using one-hot encoding and standardization within a pipeline, and model performance was assessed using stratified five-fold cross-validation. The tuned Random Forest achieved the most balanced performance, with 0.956 accuracy, 0.861 delayed-class precision, 0.805 delayed-class recall, 0.832 delayed-class F1-score, and 0.977 ROC-AUC. The tuned SVM with a threshold of 0.30 achieved higher delayed-class recall (0.857) and ROC-AUC (0.980). Feature-importance analysis indicated that fourth and fifth semester GPAs were the strongest predictors. These findings show that machine learning can support early academic intervention and data driven decision making in higher education.
Copyrights © 2026