GDG: An evolutionary oversampling framework integrating Gaussian mixture modeling and genetic algorithm

Abstract

Class imbalance induces significant bias in machine learning classifiers. While oversampling mitigates this, a critical knowledge gap exists: conventional generative methods often assume unimodal distributions and fail to address complex boundary overlap, leading to noisy, low-fidelity synthetic samples. To bridge this gap, we propose GDG, a novel framework integrating Gaussian Mixture Model (GMM) and Genetic Algorithm (GA). First, GMM clusters minority samples to accurately capture intrinsic multi-modal structures. Subsequently, an innovative global–local mechanism adaptively allocates synthetic samples based on boundary complexity, effectively minimizing overlap. Lastly, the GA performs a nonlinear search within superspheres, utilizing adaptive fitness weights to balance exploration and exploitation for high-quality generation. Extensive experiments on 21 benchmark datasets demonstrate that GDG significantly outperforms nine state-of-the-art baselines, improving average Accuracy by 1.9%, G-mean by 6.0%, and AUC by 1.2%. Rigorous non-parametric statistical analysis confirms these differences ($p=1.78\times 10^{-7}$), with post-hoc Nemenyi testing verifying that GDG achieves the superior average rank of 2.17. These findings establish GDG as a robust, statistically validated solution for tackling complex class imbalance problems.