Subgraph-Aware Graph Representation Learning for Collaborative Risk Scoring and Organized Fraud Detection

Abstract

This paper addresses the challenges of high-frequency transactions, complex transaction structures, and covert collaborative crimes in anti-money laundering and anti-fraud scenarios. It investigates a joint modeling method for transaction risk scoring and gang identification, aiming to simultaneously characterize transaction-level risk intensity and subgraph-level group associations within a unified framework. First, transaction data is abstracted into a directed transaction graph, defining risk scoring tasks for transaction nodes and gang identification tasks for suspicious transaction communities (or high-risk subgraphs). A unified embedding representation for transaction nodes and transaction subgraphs/communities is achieved through shared representation learning. Then, a graph representation module based on a graph encoder is constructed to aggregate and fuse transaction attributes and neighborhood structure information. Risk scoring heads and gang identification heads are designed separately for the shared embedding, outputting continuous risk scores and gang probability distributions. A multi-task joint objective is introduced to constrain the consistency and stability of the shared representation between the two tasks. This method can more fully utilize multi-hop fund transmission and local subgraph patterns at the transaction-chain (multi-hop) level, reducing risk fragmentation caused by relying solely on single transaction features, and capture potential collaborative structures at the subgraph/community level to assist in risk handling prioritization and clue closure. Comparative experiments verified the comprehensive advantages of the proposed joint modeling framework across multiple evaluation indicators, demonstrating its effectiveness and application value in transaction monitoring and risk governance tasks.