Contrastive Learning for Money Laundering Detection: Node-Subgraph-Node Method with Context Aggregation and Enhancement Strategy

Money laundering is the process of legitimizing funds obtained through illegal activities. It significantly impacts the global economy and fosters various criminal activities (e.g., terrorism, drug smuggling, or human trafficking). Existing money laundering detection methods face performance and robustness challenges due to low data quality and scarce labels. This paper proposes the CA-ES framework (Contrastive Pre-Trained Transformer Encoder with Context Aggregation and Enhancement Strategy). The framework employs self-supervised contrastive learning techniques to pre-train the Transformer Encoder, enabling the model to learn high-quality representations of samples without requiring labels. Specifically, we propose the Node-Subgraph-Node method, which adequately expresses the context information of each node. Additionally, we design a context aggregation and an enhancement strategy to fully mine graph knowledge, thereby generating high-quality representations that are beneficial to money laundering detection. To the best of our knowledge, this is the first proposal to apply the Contrastive Pre-Trained Transformer Encoder to anti-money laundering. Evaluation on two real Bitcoin transaction datasets shows that our proposed method outperforms seven baseline methods on multiple evaluation metrics. Under the Elliptic dataset, our money laundering detection method achieves Precision, Recall, F1-Score, and AUC-Score of 0.967, 0.875, 0.919, and 0.936, respectively. Compared to the best baseline method, Recall, F1-Score, and AUC-Score improved by 15.4%, 9.1%, and 2.0%, respectively.