ECCFD-GNN: A Novel Risk-Sensitive Graph Neural Network Model for Fraudulent Transaction Detection

The study presents the integration of machine learning techniques for detecting the credit card fraud. Its integration maintains a behavioral profile of cardholders and other parameters like location, frequency, amount etc. resulting in the timely detection of any anomaly from the normal behavior. A novel approach ECCFD-GNN (Enhanced Credit Card Fraud Detection based on Graph Neural networks) is proposed enhancing the performance of fraud detection. The various behavioral indicators taken into consideration are number of months with late payments, the frequency of low payments and the length of the account, which are further combined to a newly introduced feature “risk score”. The purpose of risk score is to increase the model’s sensitivity to the transactions having complex fraud risks. The approach uses three Graph Neural Network architectures namely KNN graph GNN, Radius Graph GNN and Feature Correlation GNN. The experiment is performed with both the optimizers Adam and RAdam. With Adam optimizers the results show that KNN graph GNN provides better performance when compared on the basis of different evaluation parameters with accuracy 85%, precision 76% recall 70% and F1-score as 73%.  The results are improved when tested with RAdam optimizers leading to increased accuracy, precision, recall and F1 score.