Weight-Based Oversampling and Overlapping Data Modification for Imbalanced Student Dropout Prediction

Abstract

Predicting student dropout in higher education is challenging due to class imbalance and class overlap, particularly in multiclass classification. This study presents a hybrid method that combines weight-based oversampling and energy-based cleaning to address class imbalance and class overlap. The method generates synthetic minority samples using neighborhood information and interpolation directions with lower overlap risk. It also adjusts the majority samples in overlapping regions by modifying their feature values toward their class center. Experiments were conducted on an Indonesian public university dataset using several machine learning classifiers and stratified 10-fold cross-validation. The proposed method achieved the best overall performance, with mGM=0.848, MAUC=0.923, and dropout recall=0.918, significantly outperforming several resampling methods according to the Wilcoxon signed-rank test. These results indicate that addressing class imbalance and class overlap can improve dropout identification in multiclass settings.