A Spatio-Temporal Attention-Based GCN for Anti-money Laundering Transaction Detection

Abstract

Money laundering is a serious financial crime that has significant implications for a country’s economy, finance, and political stability. Machine learning and deep learning methods have been used to identify instances of money laundering, with some notable successes. However, most studies focus on static structure of transaction graph, ignoring the dynamic of transactions over time. In this study, we propose a novel approach called TemporalGAT that leverages temporal and spatial attention mechanisms to improve the accuracy and efficiency of money laundering detection. Specifically, we employ multi-head attention mechanisms to perform node embedding on spatial structure of graph, extract features from transaction data, and introduce a dynamic update mechanism that enables the LSTM to adaptively update graph convolutional network parameters over time. This approach allows the model to capture dynamic changes and spatial correlations in transaction data. We evaluate the proposed method on the publicly available Elliptic dataset for node (transaction entity) classification tasks, and the experimental results demonstrate that TemporalGAT outperforms existing methods in money laundering transaction detection.

Pengwei Wang is the corresponding author. This work was partially supported by the National Natural Science Foundation of China (NSFC) under Grant 61602109, DHU Distinguished Young Professor Program under Grant LZB2019003, Shanghai Science and Technology Innovation Action Plan under Grant 22511100700, Fundamental Research Funds for the Central Universities, the Key Innovation Group of Digital Humanities Resource and Research of Shanghai Municipal Education Commission.