Graphs in Anti-Fraud and How We Apply Them

Islam Umarov1 Nov 2024

Problem

A review platform needs to remove paid and coordinated reviews. Rules and models on one review, one user or one organization see only that entity. Fraudsters adapt to the rules, and a single account can look normal. The signal sits in the connections: one account reviews both a school canteen and a Bentley showroom, or a group of accounts praises one organization and attacks all its neighbors. The task is also unusual for ML, because the goal is to find fraud no label describes yet, not to fit the known cases.

Example from the article: an account that leaves reviews for a school canteen and a Bentley showroom at the same time, probably a schoolchild selling reviews.

Solution in brief

Treat reviews as a graph and add graph methods to the existing rules and models. REV2 propagates user, review and organization trust scores and tightens current rules. GraphSAGE-style aggregation of neighbor features finds new fraud. Temporal Graph Networks add event time.

Questions this case answers
  • How to detect coordinated manipulation by scoring each entity through its connections in a bipartite interaction graph, where single-entity rules see nothing
  • How to add neighbor-aggregated graph features to an existing rule and gradient-boosting detection stack when the target behavior has no labels yet
  • How to catch review manipulation that per-review and per-user rules miss
  • How to add graph signals to an existing rule-plus-ML anti-fraud stack

Tools and methods

  • REV2 — iterative scores for user fairness, review reliability and organization quality
  • GraphSAGE — encoder, neighbor convolution, decoder, per-node fraud prediction
  • Neighborhood Feature Aggregation + CatBoost — min, max and avg of neighbor features as extra model inputs
  • Temporal Graph Networks (TGN) — node memory updated by each new edge
Continue reading — it's free

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