Complex network analysis of bank transaction networks and unsupervised graph-based detection of money laundering

Abstract

Financial crime, particularly money laundering, poses a significant threat to the global financial system, with illicit funds flowing into various criminal activities. While much traditional research has focused on identifying fraudulent entities in bank transactions, it often overlooks the fact that money laundering activities involve the participation of multiple accounts, forming irregular transaction patterns. Unlike traditional methods, a graph-based money laundering detection approach focuses on the connections among accounts to detect anomalous behavior.

This research conducted a network analysis on two empirical bank transaction networks to explore their structural characteristics. The analysis revealed that these networks can be considered multi-centered and exhibit scale-free, small-world, and preferential attachment properties. Additionally, the study provides a framework for reconstructing networks using a modified random walk subgraph sampling method, effectively preserving the topology of the original networks.

The research further examines several money laundering related patterns in networks and discuss various types of features that can be extracted from networks for money laundering detection. Leveraging these extracted features, an unsupervised anomaly detection approach using the IsolationForest model, was applied to identify outliers that may indicate suspicious activities. Our findings reveal that features generated through graph random walks are effective in detecting fraud patterns such as cliques and cycles. The results demonstrate the potential of using graph-based features for anomaly detection without relying on personal data, offering a promising supplementary tool to existing anti-money laundering methods.