Fraud Detection with Graph Convolutional Networks - An evaluation on the classification performance of GCN when applied on simulated bank transaction logs

Abstract

Money laundering is the process for making proceeds from illegal businesses seem legal. The “history” of the money is “laundered away” by being forwarded between links in a chain – or rather nodes in a complex network of accounts and agents – so called money mules. This to cover the origin, and the true owner, of the money or valuables. The United Nations Office on Drugs and Crime estimates that approximately of 2-5% of global GDP is laundered every year. This is money invested in further illegal businesses, such as human trafficking, fraud, corruption, and terrorism. Through Anti-Money Laundering initiatives and legislation money laundering can be suppressed. Banks and other financial institutions are by law compelled to inquire about and report suspicious transactions. However, this is an expensive and cumbersome task for institutions to take on since it requires a lot of manual labour. This while a lot of money mule networks remain undetected. Previous studies show that the application of Graph Convolutional Network models perform well on multiple class classification problems such as social networks, citations, and fraud detection. The aim of this study is to evaluate the performance of GCN when applied on an imbalanced, binary classification problem. The data in question, AMLSim, simulates bank transaction logs where agents in the role of bank customers perform in a typical fraudulent behaviour. However, results are inconsistent and the performance of the GCN when applied on AMLSim is dissatisfactory.