FraudMemory: Explainable Memory-Enhanced Sequential Neural Networks for Financial Fraud Detection

Abstract

The rapid development of electronic financial services brings significant convenience to our daily life. However, it also offers criminals the opportunity to exploit financial systems to do fraudulent transactions. Previous studies on fraud detection only deal with single type transactions and cannot adapt well to evolving environment in reality. In addition, their black box models pay less attention on the interpretability of fraud detection results. Here we propose a novel fraud detection algorithm called FraudMemory. It adopts state-of-art feature representation methods to better depict users and logs with multiple types in financial systems. Our model innovatively uses sequential model to capture the sequential patterns of each transaction and leverages memory networks to improve both the performance and interpretability. Also, with the incorporation of memory components, FraudMemory possesses high adaptability to the existence of concept drift. The empirical study proves that our model is a potential tool for financial fraud detection.