Experimental Evaluation of Generative Adversarial Network for Addressing Class Imbalance in Machine Learning

Abstract

Class imbalance poses a significant challenge in machine learning, especially in critical domains like fraud detection and healthcare. The dominance of majority classes often overshadows minority classes, leading to models that inadequately recognize rare but pivotal events, such as fraudulent transactions. This imbalance can compromise predictive accuracy and fairness. Traditional methods, including resampling techniques and ensemble methods, often suffer from overfitting and inadequate representation. This study explores the use of Generative Adversarial Networks (GANs) to enhance predictive performance on an imbalanced dataset of credit card transactions, comprising 95% legitimate and 5% fraudulent transactions. The GAN architecture consists of a generator producing synthetic samples of the minority class and a discriminator distinguishing between real and synthetic data. Experimental results indicate models augmented with GANs achieve high accuracy, with Random Forest models reaching 99.98%, Gradient Boosting models achieving 99.99%, and Decision Tree models obtaining 99.82%. These findings underline GANs' effectiveness in addressing class imbalance, enhancing predictive performance for minority classes and providing reliable results in practical applications.