Improving Student Graduation Timeliness Prediction Using SMOTE and Ensemble Learning with Stacking and GridSearchCV Optimization

Introduction: Timely graduation is a key performance indicator in higher education. This study aims to improve the accuracy of predicting student graduation timeliness using ensemble machine learning techniques combined with SMOTE and hyperparameter optimization. Methods: This is a quantitative predictive study. The population includes students and alumni of Universitas Islam Riau. A sample of 160 respondents was obtained via purposive sampling. Data were collected using structured questionnaires encompassing academic variables (e.g., GPA, credits, attendance) and non-academic variables (e.g., stress, social support, extracurricular activity). After preprocessing and label encoding, SMOTE was applied to balance class distribution. Several classifiers (Naïve Bayes, SVM, Decision Tree, KNN) were tested, with ensemble learning (voting and stacking) implemented and optimized using GridSearchCV. Results: The stacking ensemble model optimized with GridSearchCV achieved the highest performance with an accuracy of 99.37%, precision and recall above 0.99, and minimal misclassification. This outperformed individual models and previous approaches in the literature. Conclusions: The integration of SMOTE, ensemble methods, and GridSearchCV significantly enhances predictive accuracy for student graduation timeliness. The resulting model provides a robust framework for academic risk detection and early intervention.