Financial fraudsters use Gen AI, digital channels, global networks, and synthetic identities making it complex to identify the fraudulent activities. Traditional rule-based systems relying on traditional methods do not identify frauds which use multi-step transaction routing with multiple institutions and across borders. Graph database using Labelled Property Graphs, represents customers, accounts, and transactions as interconnected nodes and edges. By ingesting live transaction data, they apply pattern-matching and community-detection to expose suspicious subgraphs. Money-laundering rings or collusive clusters—and let investigators trace multi-hop links to “hub” accounts with clear visual audit trails. Machine learning models trained on vast historical datasets, using supervised classifiers (e.g., gradient boosting) and unsupervised anomaly detectors. Features like transaction amounts, geolocation consistency, device fingerprints, and temporal sequences feed these models, while recurrent architectures capture evolving fraud tactics. Yet they often suffer from concept drift, require extensive labelled data, underperform on imbalanced cases, and behave as opaque black boxes, generating false positives and hampering trust. A hybrid framework combines relational graph insights with statistical scoring, boosting detection accuracy, reducing false alarms, and enhancing investigators’ confidence in fraud detection and prevention.
Babu et al. (Thu,) studied this question.