Graph Neural Networks для антифрода

Graph Neural Networks (GNN) — прорыв для задач с relational data: fraud rings, collusion, money laundering.

Motivation

Classical ML: features per entity. Не capturing relationships. Пример: individual user выглядит normal, но связан с 10 fraud accounts через device. Graph structure — critical signal. GNN natively handles эти relationships.

Основные архитектуры

GCN (Graph Convolutional Networks): базовая message passing. GAT (Graph Attention Networks): learned importance neighbors. GraphSAGE: sampling для scalability на больших графах. Heterogeneous GNNs: разные типы узлов и рёбер.

Fraud applications

Community detection: identify fraud rings (10–1000 связанных accounts). Ranking: fraud score с propagation через связи. Anomaly detection в графе. Link prediction: identify hidden connections. Money laundering trails: unique patterns в transaction graphs.

Реализация

PyTorch Geometric, DGL: основные frameworks. Neo4j для storage графов. Feature engineering: node features + edge features. Sampling strategies для больших графов (billions of edges). Training: 10x compute vs XGBoost, но 20–40% AUC lift на fraud rings.