Money Laundering Detection Using A Transaction-Based Graph Learning Approach

Money laundering poses a pervasive threat to the stability and integrity of global financial systems. Since traditional anti-money laundering (AML) methods predominantly rely on rule-based systems and statistical approaches, it has limitations to capture the intricate and interconnected relationships that is inherent in money laundering networks. In response to this challenge, this paper proposes an innovative approach to enhance money laundering detection through transactions. We begin by constructing network graphs from a large dataset of bank transactions. Drawing insights from language modeling and supervised learning, we transform these graphs into directed node representations that effectively encode these intricate relationships and community structures within the transaction network. Subsequently, we utilize Random Forest (RF) to predict suspicious behavior associated with money laundering. Additionally, we address the specific challenges posed by highly imbalanced classes in the context of money laundering detection through both oversampling and undersampling experiments to overcome these challenges. The predictive performance of the RF model with oversampling yielded an accuracy of 86%, whereas when undersampling was applied, the accuracy increased to 92%.