Benchmarking state-of-the-art imbalanced data learning approaches for credit scoring

The goal of credit scoring is to identify abnormalities, aiding decision making and maintaining the order of financial transactions. Due to the small number of default records, one inevitably faces a class imbalance problem when handling financial data. The class imbalance problem has received a lot of attention because of the economic loss that can occur when one fails to accurately identify default samples. To solve this problem, there are various classic and mature approaches to learning imbalanced data, including resampling approaches, cost-sensitive strategies, and so on. Especially in recent years, generative adversarial networks (GANs) have attracted the attention of researchers to explore these networks’ effects as imbalanced data learning tools. However, no attention has been paid to the systematic scoring and comparison of these traditional and state-of-the-art imbalanced data learning approaches in relation to credit scoring. Therefore, choosing several related datasets, we compare the performance of the traditional approaches and GANs in solving the class imbalance problem of credit scoring; at the same time, with the help of benchmark analysis, we provide some suggestions for relevant research.