BSGAN: A Novel Oversampling Technique for Imbalanced Pattern Recognitions

TL;DR

BSGAN introduces a hybrid oversampling method combining borderline SMOTE and GANs to generate diverse, Gaussian-distributed data, effectively addressing class imbalance in pattern recognition tasks.

Contribution

This work presents BSGAN, a novel oversampling technique that improves diversity and distribution of synthetic samples beyond existing borderline SMOTE methods.

Findings

BSGAN outperforms existing oversampling techniques on multiple datasets.

Generated data from BSGAN follow Gaussian distributions.

BSGAN enhances data diversity after oversampling.

Abstract

Class imbalanced problems (CIP) are one of the potential challenges in developing unbiased Machine Learning (ML) models for predictions. CIP occurs when data samples are not equally distributed between the two or multiple classes. Borderline-Synthetic Minority Oversampling Techniques (SMOTE) is one of the approaches that has been used to balance the imbalance data by oversampling the minor (limited) samples. One of the potential drawbacks of existing Borderline-SMOTE is that it focuses on the data samples that lay at the border point and gives more attention to the extreme observations, ultimately limiting the creation of more diverse data after oversampling, and that is the almost scenario for the most of the borderline-SMOTE based oversampling strategies. As an effect, marginalization occurs after oversampling. To address these issues, in this work, we propose a hybrid oversampling technique by combining the power of borderline SMOTE and Generative Adversarial Network to generate more diverse data that follow Gaussian distributions. We named it BSGAN and tested it on four highly imbalanced datasets: Ecoli, Wine quality, Yeast, and Abalone. Our preliminary computational results reveal that BSGAN outperformed existing borderline SMOTE and GAN-based oversampling techniques and created a more diverse dataset that follows normal distribution after oversampling effect.