Can Generative Adversarial Networks Help Us Fight Financial Fraud?

Transactional fraud datasets exhibit extreme class imbalance. Learners cannot make accurate generalizations without sufficient data. Researchers can account for imbalance at the data level, algorithmic level or both. This paper focuses on techniques at the data level. We evaluate the evidence of the optimal technique and potential enhancements. Global fraud losses totalled more than 80 % of the UK’s GDP in 2019. The improvement of preprocessing is inherently valuable in fighting these losses. Synthetic minority oversampling technique (SMOTE) and extensions of SMOTE are currently the most common preprocessing strategies. SMOTE oversamples the minority classes by randomly generating a point between a minority instance and its nearest neighbour. Recent papers adopt generative adversarial networks (GAN) for data synthetic creation. Since 2014 there had been several GAN extensions, from improved training mechanisms to frameworks specifically for tabular data. The primary aim of the research is to understand the benefits of GANs built specifically for tabular data on supervised classifiers performance. We determine if this framework will outperform traditional methods and more common GAN frameworks. Secondly, we propose a framework that allows individuals to test the impact of imbalance ratios on classifier performance. Finally, we investigate the use of clustering and determine if this information can help GANs create better synthetic information. We explore this in the context of commonly used supervised classifiers and ensemble methods