PageRank-Based Unsupervised Deep Vertex Representations for Anti-Money Laundering Detection

Abstract

Anti-money laundering (AML) refers to a global framework of laws, regulations, and protocols designed to trace and expose illicit funds that have been concealed to appear legitimate. Enforcing anti-money laundering measures demands rigorous monitoring of financial transactions to detect potentially unlawful activities. However, conventional rule-based AML systems, while compliant with regulatory standards, generate excessive false positives that burden analysts and obscure subtle laundering patterns. To address these limitations, advanced machine learning techniques, particularly those based on graph structures like Graph Neural Networks (GNNs), offer a promising alternative. In this context, we introduce a novel method, referred to as AML PD (Anti-Money Laundering with Personalized Diffusion), which utilizes Personalized PageRank and diffusion mechanisms to generate unsupervised node embeddings. This technique learns inductive graph embeddings for AML detection by taking into account the directionality of graph edges and incorporating both node and edge features, along with the graph's structural information. AML PD derives a node's local representation by fusing diffusion processes with Personalized PageRank scores. This fusion captures critical information relevant to AML when transformed into a lower-dimensional embedding space. These embeddings are then used by a classifier to identify potential instances of money laundering. Our method is not only scalable to large datasets but also enables deeper analysis of the learned embeddings. Experimental results indicate that our method surpasses existing baseline models, particularly when positional encodings are integrated. A core objective of our work is to enhance the efficiency of AML investigations by offering AI-powered insights. As such, AML PD demonstrates strong potential in improving the performance of GNN-based AML detection systems.