Synthetic Minority Over-sampling Technique for Student Performance Prediction: A Comparative Analysis of Ensemble and Linear Models
Khaled Mahmud Sujon, rohayanti binti hassan, Nusrat Jahan · 2024
This study explores the impact of the Synthetic Minority Oversampling Technique (SMOTE) on predictive accuracy in student performance prediction using both linear (Logistic Regression, LR) and ensemble (Random Forest, RF) models. We compare the performance of these models, using a real-time dataset from Universiti Teknologi Malaysia (UTM). While previous studies have examined these models separately, this research uniquely investigates the simultaneous application of SMOTE to both. Our results show that LR’s accuracy improves from 91% to 95% with SMOTE, while RF maintains a consistent 98% accuracy. These findings demonstrate that SMOTE significantly enhances the performance of linear models in handling imbalanced data, providing useful insights for curriculum development in higher education. This research contributes to the literature by guiding higher education institutions in selecting the most effective predictive models.