International Conference on Computer and Knowledge Engineering

Home / 15th International Conference on Computer and Knowledge Engineering

SUBoost: A Novel Boosting-Based Selective Undersampling for handling Imbalanced Data

Authors :

Nima Rasi Baghmishe¹ Jafar Tanha² Ehsan Roshan³

1- Faculty of Electrical and Computer Engineering University of Tabriz 2- Faculty of Electrical and Computer Engineering University of Tabriz 3- Faculty of Electrical and Computer Engineering University of Tabriz

Keywords :

AdaBoost،Ensemble methods،Undersampling،Imbalanced datasets،Classification

Abstract :

Abstract—One of the persistent challenges in machine learning is dealing with class imbalance, where one class (the majority class) contains significantly more samples than the other (the minority class). Although ensemble methods like AdaBoost have proven effective across various types of data, their performance tends to suffer on imbalanced datasets, as they often favor the majority class. In this study, we introduce a new variant of the AdaBoost algorithm, named SUBoost, designed specifically to address this issue. Unlike conventional random undersampling techniques, SUBoost selectively reduces the majority class based on classifier performance, leading to a more effective balance between classes. We assess the performance of SUBoost against seven established ensemble methods across 17 imbalanced datasets, using AUC and G-Mean as evaluation metrics. The experimental results demonstrate that SUBoost consistently outperforms existing methods and offers a notable improvement in managing class imbalance.