AWGAN: An adaptive weighting GAN approach for oversampling imbalanced datasets

Oversampling is a widely employed technique for addressing imbalanced datasets, facing challenges like class overlaps, intra-class imbalance, and noise. In this paper, we introduce an adaptive weighted oversampling algorithm grounded in generative adversarial networks, which we term AWGAN. To begin, our method computes the local and global densities for each instance, confirming its distribution within its local neighborhood, thereby enabling accurate identification and elimination of noisy instances. Subsequently, we devise a weight calculation strategy based on boundary division. Minority class instances are classified into safe and boundary instances, and weights are calculated based on the density of each instance and its distance from the surrounding instances, assigning different weights to overlapping and non -overlapping regions, and sparse and dense region instances, in order to solve the problems of class overlap and intraclass imbalance. Finally, GAN is used to construct a balanced dataset by adaptively generating minority class instances that match the real data distribution based on the weights. We evaluate AWGAN against six traditional oversampling methods and five GAN-based oversampling methods. The experimental results demonstrate that AWGAN significantly enhances classifier performance, as evident in its F1 -Score, AUC, G -mean, and MCC on 21 diverse datasets.