Financial Fraud Detection with Improved Neural Arithmetic Logic Units

Domain specific neural network architectures have shown to improve the performance of various machine learning tasks by large margin. Financial fraud detection is such an application domain where mathematical relationships are inherently present in the data. However, this domain hasn’t attracted much attention for deep learning and the design of specific neural network architectures yet. In this work, we propose a neural network architecture which incorporates recently proposed Improved Neural Arithmetic Logic Units. These units are capable of modelling mathematical relationships implicitly within a neural network. Further, inspired by a real-world credit payment application, we construct a synthetic benchmark dataset, which reflects the problem setting of automatically capturing such mathematical relations within the data. Our novel network architecture is evaluated on two real-world and two synthetic financial fraud datasets for different network parameters. We compare our proposed model with several well-established classification approaches. The results show that the proposed model is able to improve the performance of neural networks. Further, the proposed model performs among the best approaches for each dataset.