SafeVAE-GAN: A Novel Re-sampling Approach for Multi-Class Imbalanced Data

Abstract

Multi-class imbalanced datasets are ubiquitous in domains like medical diagnostics, fraud detection, and learning performance classification, where minority classes are critical but underrepresented, and class overlap introduces noise and ambiguous boundaries. Prior research has explored oversampling and undersampling techniques, generative approaches, and hybrid sampling methods to balance multi-class datasets, aiming to enhance classifier performance across imbalanced and overlapping classes. These methods often produce noisy samples near class boundaries, fail to preserve class-specific characteristics, or struggle with high-dimensional data and severe imbalances, particularly in multi-class settings, resulting in poor performance on minority classes. In this article, we propose SafeVAE-GAN, a resampling framework that integrates a conditional β-Variational Autoencoder (β-VAE) with Wasserstein GAN to generate high-quality minority samples. Our framework includes two complementary variants: SafeVAE-GAN O , an advanced oversampling method using dual Safe Region Constraints; and SafeVAE-GAN H , a hybrid variant that further performs targeted undersampling of majority samples near ambiguous boundaries based on k-NN heterogeneity. Across 26 benchmark datasets, the proposed framework demonstrates competitive performance and consistent improvements in Macro-F1, Macro G-mean, mGM and Macro-MCC over strong baselines. Overall, SafeVAE-GAN provides a robust and scalable solution for multi-class imbalanced learning across diverse domains.