Graph-Based Deep Learning for Financial Crime Detection: A Review of Transaction Networks, Fraud Patterns, and Explainable AI.

Financial fraud and money laundering now operate through networks of accounts, devices, merchants, and identities, and rule-based systems and conventional machine learning models that score transactions one at a time often miss them. This narrative review examines how Graph Neural Networks (GNNs) detect suspicious connections among entities in transaction networks. It explains how financial data is represented as graphs and how message passing works, then reviews applications to money laundering and crime rings, synthetic identities and account takeover, and merchant fraud. It compares GNNs with conventional methods on accuracy, class imbalance, scalability, and interpretability. The reviewed studies report that GNNs can find structural patterns, such as hubs, cycles, and dense closed loops, that attribute-based models miss, and that sampling and label-balancing techniques help with imbalanced fraud data. The evidence is not one-sided: on the Elliptic Bitcoin benchmark, a random forest outperformed a graph convolutional network, and GNNs can struggle when labels are heavily skewed. Computational cost, camouflaged fraudsters, and the difficulty of explaining graph-based decisions to regulators remain open problems. The review concludes that GNNs should be evaluated against strong conventional baselines and paired with explanation tools, federated learning, and human oversight. The article follows no formal search protocol.