Addressing Dataset Imbalance with Modified Generative Adversarial Networks

Abstract

Class imbalance in datasets remains a major problem in machine learning; most algorithms produce a biased model that generalizes poorly on the minority classes. This becomes an important factor in applications where minority Class samples are very important, such as medical diagnosis, accounting fraud detection, financial analysis, and so on. This paper proposes a new method to overcome the problem of class imbalance by using Generative Adversarial Networks (GANs). Specifically, the proposed method aims at obtaining a set of synthetic instances for the minority classes, which helps to balance the training set and improve classifiers’ performance. It introduces a modified Generative Adversarial Network (MGAN), which generates synthetic images of the minority class (128 × 128 × 3 pixels), thereby enhancing classifier resilience. Empirical evaluation indicates that MGAN surpasses conventional data balancing methods in medical imaging contexts. The effectiveness of this image filtering approach regarding lesion detection is experimentally confirmed when using two medical datasets of endoscopic and pathological images. The outcomes represent a high performance of the variant classifiers, which outperforms the conventional approaches when using the altered MGAN.