Authors: Assistant Professor Anitha K, Assistant Professor Saranya B
Abstract: With the rise in adoption of real-time payment systems such as UPI, Pix, and FedNow comes a unique set of problems related to detecting fraud within strict latency requirements, and where fraudsters are acting in a networked manner. The conventional methods of rule-based and tabular machine learning work by treating transactions as individual entities and cannot account for the relationships between them and how it plays into the nature of today’s fraud schemes such as mule networks and authorized push payment scams. In this paper, we present a novel approach using graph neural networks to solve the problem of real-time fraud detection by representing the payment network as a dynamic and heterogeneous graph consisting of accounts, devices, merchants, and transaction flows. Our approach combines GraphSAGE for inductive representation learning and LightGBM for tabular feature extraction and can perform sub-millisecond predictions while being able to update in a streaming fashion. Our experiments on real transaction data show improvements over baseline methods in terms of fraud detection recall from 0.71 to 0.84 with the false positives being less than 5%.
