Handling Class Imbalance in Student Success Prediction Using Machine Learning: A Comparison of SMOTE and SMOTETomek

Predictive models for student academic success are often hindered by class imbalance, particularly in identifying students at risk of dropping out. This study evaluates nine machine learning algorithms under three data balancing scenarios: unbalanced, SMOTE, and SMOTETomek. Using a dataset of 4,424 records from the Polytechnic Institute of Portalegre, Portugal, we found that Random Forest combined with SMOTETomek achieved the best performance, attaining 77.62% accuracy and a 50.31% recall for the Dropout class. By effectively reducing false negatives without sacrificing accuracy, this approach offers a practical early warning solution for educational institutions.