Build a Fraud-Ring Graph Pipeline Without Blowing Up on Hubs

A payments company has 5,000,000 accounts, 3,000,000 devices and 1.2 billion transfers. The fraud team wants to catch rings (groups of accounts that share devices and pay each other) earlier. An analyst suggests building an account-to-account graph by connecting any two accounts that ever used the same device, then running a GNN.

1. Why is that projection dangerous? Quantify with a device used by 1,000 accounts and a typical device used by 10.
2. Propose a graph schema and a feature pipeline that avoids the problem.
3. Explain how rings show up in classic features and name the features you would compute.
4. Describe how the output reaches the decision system and what you would monitor.