Emerging SMOTE and GAN Variants for Data Augmentation in Imbalance Machine Learning Tasks: A Review
Amadi Gabriel Udu, M. K. Salman, Maryam Khaksar Ghalati, Andrea Lecchini‐Visintini, D. R. Siddle, Hongbiao Dong
Abstract
Class imbalance is a pervasive challenge in real-world machine learning (ML) applications, where the minority class, often the class of interest, is significantly underrepresented. This imbalance can negatively affect model performance, lead to misleading evaluation metrics, and introduce validation challenges. Two prominent data-augmentation techniques to address class imbalance are the Synthetic Minority Oversampling Technique (SMOTE) and Generative Adversarial Networks (GAN). However, both techniques have their inherent limitations, motivating the emergence of novel variants designed to overcome these challenges. While previous reviews have primarily focused on specific domains, traditional methodologies, or broad strategy overviews, this reviewpresents a unified taxonomy that captures the causes, types, and implications of class imbalance across diverse ML tasks. It further explores emerging trends in SMOTE and GAN applications, limitations, and hybrid adaptations. By categorising imbalance types and examining models, metrics, datasets, and comparative approaches, this review provides actionable insights and future research directions for practitioners and researchers addressing class imbalance in real-world ML tasks.