Graph-Driven Strategies for Combating Money Laundering

Abstract

This research uses graph-based machine learning models to combat money laundering. The study aims to create a robust anti-money laundering (AML) framework that can distinguish between legal and illegal financial transactions. Key aspects of the research include: Extending traditional graph neural networks (GNNs) to process directed multigraphs using techniques like reverse message passing, multigraph port numbering, and ego IDs. Enriching machine learning models with graph-based features using TigerGraph, a native graph database. Using a synthetic dataset that simulates a financial transaction network, addressing issues of privacy and limited real-world data. The dataset is split 60-20-20 for training, validation, and testing The results show that the XGBoost model augmented with graph features performed the best, achieving an AUC of 87.26% and an F1-score of 64.77%. The GIN model, enhanced with port numbering, ego IDs, and reverse message passing, also performed well, with an AUC of 83.27% and an F1-score of 59.14%. Key graph features include 'scatter_gather', 'fan_out', and 'weighted_pagerank'. Our research seeks to empower financial institutions to improve AML compliance and detection accuracy. Future work may integrate graph embedding algorithms and evaluate the models on larger datasets