Unsupervised graph anomaly detection - Catching new fraudulent behaviours

Rizal Fathony2 Aug 2023

Problem

Fraudsters keep changing their modus operandi (MO), so a supervised model trained on historical labels cannot recognise a pattern it has never seen. A food and grocery delivery platform needs a signal that flags such new MOs before labels exist, both for suspicious users and for suspicious orders. The signal must cover consumers, merchants and the orders between them.

Example from the article: Grab modelled GrabFood and GrabMart consumer-merchant orders as a graph, and the anomaly model led to the discovery of new fraud MOs (the article names none).

Solution in brief

Build a bipartite graph of consumers and merchants with orders as edges, and train a graph autoencoder (GraphBEAN) that reconstructs node and edge attributes and the edge structure. Rare behaviour reconstructs badly, so reconstruction error becomes an anomaly score for each edge and each node. A heuristic tagger labels the fraud type, and the scores go to analysts and to automatic actions.

Questions this case answers
  • How to detect previously unseen adversarial behaviour in a two-sided platform when no labels exist for the new patterns
  • How to score both entities and their interactions for anomalies when each entity group has its own attributes and each interaction carries attributes too
  • How to detect new fraud patterns that are absent from historical labeled data?
  • How to score both users and their interactions for anomalies on a consumer-merchant graph without labels?

Tools and methods

  • GraphBEAN — graph autoencoder that takes node and edge attributes and outputs node-level and edge-level scores
  • Graph convolution layers — encoder and feature decoder that pass information between nodes and edges
  • Reconstruction-based anomaly score — error of rebuilding features and structure, aggregated per node
  • Fraud type tagger — heuristic rules that attach tags such as promo abuse or possible collusion to detected anomalies
Continue reading — it's free

Google or a one-time email code. No password.