Explainable Financial Fraud Detection Using Graph Neural Networks: A Comparative Study

Abstract The financial fraud detection problem is extremely complicated and difficult because of the relationships between transactions in financial systems and the extreme imbalance of fraudulent and legitimate transactions. Traditional machine learning methods might not be suitable for modeling the relational nature of transaction networks, which is the reason why graph neural networks (GNNs) are employed in graph-based fraud detection. This research evaluates the performance of three different GNN architectures: Graph Convolutional Networks (GCN), GraphSAGE, and Graph Attention Networks (GAT) in the field of financial fraud detection using Elliptic's Bitcoin transaction dataset. The dataset is a temporal graph of bitcoin transactions, allowing to analyze transaction attributes and relationships. The study creates a comparative experimental setup for testing the three architectures according to classification metrics, focusing on the precision, recall and F1 score of fraud-class accuracy together with overall accuracy. The interpretability problem of graph-based predictions is overcome by plugging GNNExplainer into a unified explanation pipeline to locate the structure of subgraphs and the features of nodes that are relevant to each prediction from the model. This framework brings together prediction, explanation, and visualization at the instance level, to facilitate a discussion of model performance and interpretability. The goals of the study are to describe the strengths and weaknesses of the architectures evaluated and to evaluate the usefulness of graph-based explanations for financial fraud detection. Experimental conclusions will be verified based on the results of the model and explanation analyses.