Gang Prediction in Transaction Graphs for Anti Money Laundering Detection

In this thesis, we propose an iterative graph coarsening-learning framework to detect money laundering gangs in bank transaction networks. Instead of classifying individual accounts, our goal is to identify groups of accounts that act together to hide illicit activities. We begin by providing a mathematical definition of gangs in the context of transaction graphs, modeling them as dense and highly interactive subgraphs that share structural and behavioral similarities, such as transaction patterns. The proposed method progressively reduces the complexity of the transaction graph through embedding-based coarsening, where nodes with similar structural and semantic patterns are merged into super-nodes representing potential groups. This process preserves the most relevant information while reducing noise and computational cost and also helping the model to generalize over different nodes and structures. After each coarsening step, we apply supervised learning on the reduced graph to detect suspicious groups. To guide the training, we design a coarsening-aware loss function that combines classification objectives with a consistency term linking each super-node to its constituent accounts. This encourages coherent group representations and improves generalization across different graph scales. We validate our approach on both synthetic and real-world bank transaction datasets and compare it against state-of-the-art graph neural network and semi-supervised embedding methods.