Unrolled GAN-based Oversampling of Credit Card Dataset for Fraud Detection
Jing-Zhong Wang, Lin Yao
Abstract
The excellent performance of most classification algorithms is based on the balance of classes. However, in anomaly detection, the classes are mostly biased. The performance of traditional machine learning algorithms applied to anomaly detection of ten fails to achieve the target effect. Considering the data source, oversampling method addresses class imbalance from data source. In light of the capability of capturing the original sample data distribution, Generative Adversarial Networks offer an inspiring oversampling solution. In this research we demonstrate the applicability of an oversampling method based on Unrolled GAN with credit card data sets. We contrast that method with traditional oversampling methods. Empirical results show the capacity of Unrolled GAN-based oversampling.