Evidential Generative Adversarial Networks for Handling Imbalanced Learning

The predictive performance of machine learning models tends to deteriorate in the presence of class imbalance. Multiple strategies have been proposed to address this issue. A popular strategy consists of oversampling the minority class. Classic approaches such as SMOTE utilize techniques like nearest neighbor search and linear interpolation, which can pose difficulties when dealing with datasets that have a large number of dimensions and intricate data distributions. As a way to create synthetic examples in the minority class, Generative Adversarial Networks (GANs) have been suggested as an alternative technique due to their ability to simulate complex data distributions. However, most GAN-based oversampling methods tend to ignore data uncertainty. In this paper, we propose a novel GAN-based oversampling method using evidence theory. An auxiliary evidential classifier is incorporated in the GAN architecture in order to guide the training process of the generative model. The objective is to push GAN to generate minority objects at the borderline of the minority class, near difficult-to-classify objects. Through extensive analysis, we demonstrate that the proposed approach provides better performance, compared to other popular methods.