Financial Fraud Identification Using Graph Neural Network And LSTM With Autoencoder-Based Data Refinement

Abstract

Credit card fraud is a major issue globally, with a loss of millions of dollars each year. Due to the evolution of different fraudsters and high data volume, detecting fraudulent credit card transactions is a challenging task. Due to the availability of transaction data with high imbalance, developing an efficient deep learning approach-based fraud transaction detection system is still challenging. Also, extracting significant descriptors from the transaction data determines the performance of the classifier. Therefore, the paper proposes a generative adversarial network (GAN) and an auto-encoder (AE) based transaction data generation system to train the proposed fraud transaction detection approach. The fraud transaction detection system was constructed using a graph neural network (GNN) and long short-term memory (LSTM) that extracts significant descriptors from the transaction data to distinguish fraud and non-fraud transactions. The performance of the proposed fraud detection system was evaluated using the European cardholder 2013 (ECH-2013) dataset and the PaySim dataset with scales such as F1-score, recall, precision, and accuracy. The proposed hybrid GNN-LSTM approach yields an accuracy of 97.91% and 98.09% when evaluated using the ECH-2013 and PaySim datasets, respectively.