Subgraph anomaly detection in financial transaction networks

Abstract

Effective anomaly detection is crucial for the success of many AI-based solutions in the financial domain, including e.g. fraud detection and risk modeling. Identifying anomaly from financial transaction networks is one of the challenging tasks that can be cast as a special instance of anomaly detection in networks. Existing methods typically attempt to detect only node-level anomalies, and assume prior knowledge to extract representative features for identifying anomalies. However, there exist collective fraudulent behaviors at the level of subgraphs rather than individual node. A ring structure for money laundering and a tree structure for pyramid schemes would be common examples. Also, in practice it is difficult to decide which features are more representative beforehand. In this paper, we introduce SADE (Subgraph Anomaly DEtection) framework that addresses these needs. SADE consists of two steps: 1) role-guided subgraph embedding, and 2) subgraph anomaly detection. Our approach for learning the subgraph embeddings allows to preserve both the local structure of subgraphs and the global structure of entire network by making use of global roles and local connections of nodes. The learnt representation allows effective use of the state of art anomaly detection approaches. Our extensive experiments on synthetic and real-world financial transaction networks demonstrate the effectiveness of SADE in learning subgraph embeddings without requiring any prior knowledge and detecting anomalous subgraphs.