A Dynamic Graph Convolutional Network for Anti-money Laundering

Abstract

Anti-money laundering (AML) is essential for safeguarding financial systems. One critical way is to monitor the tremendous daily transaction records to filter out suspicious transactions or accounts, which is time consuming and requires rich experience and expert knowledge to construct filtering rules. Deep learning methods have been used to model the transaction data by graph neural networks, and achieved promising performance in AML. However, the existing methods lack efficient modeling of the transaction time stamps, which provide important discriminative features to efficiently recognize the accounts participating money laundering. In this paper, we propose a dynamic graph attention (DynGAT) network for detecting suspicious accounts or users, which are involved in illicit transactions. The daily transaction records are naturally constructed as graphs, by considering the accounts as nodes and the transaction relationship as edges. To take the time stamps into account, we construct one transaction graph from the records within every time interval, and obtain a temporal sequence of transaction graphs. For every graph in the sequence, we not only compute the node embeddings by a vanilla graph attention network, but also explicitly develop a time embedding via a position-encoding block. Our method further captures the dynamics of the graph sequence through a multi-head self attention block on the sequence of concatenations of node embeddings and time embeddings. Moreover, we train the model by a weighted cross entropy loss function to tackle the sample imbalance problem. Experiments demonstrate that our method outperforms the existing ones in AML task.