Problem
Banks must run predefined AML rules, and by law an analyst has to investigate every alert. The article estimates the false positive rate of such rule systems at over 95%. Each alert leaves the analyst a long transaction list and only category aggregates to find the suspicious part. Supervised models are hard to train because reviews are slow and truly suspicious cases are rare.
Illustrative example: an alert fires on a customer, and the analyst must pick which transactions deserve a closer look and whether behavior changed recently.
Solution in brief
LaundroGraph builds a bipartite customer-transaction graph and trains a graph attention network to predict whether a given customer-transaction edge is real. No labels are needed. The learned embeddings give a per-transaction anomaly score, clusters of similar transactions, and a measure of how far a customer's behavior moved between time windows. All three are shown to the analyst during the investigation.
Questions this case answers
- How to flag unusual events for a human reviewer when almost no labels exist and the positive class is extremely rare
- How to measure how much an entity's behavior has shifted between time windows using learned representations instead of hand-built aggregates
- How to give compliance analysts a starting point in a pile of transactions behind an AML alert without labeled cases?
- How to score each transaction as typical or unusual for the customer who made it?
Tools and methods
- Bipartite customer-transaction graph — directed graph where customers link to their transactions and transactions link to the customers involved
- Link prediction with binary cross-entropy — self-supervised training objective
- Graph attention network (GAT) — encoder that weighs neighbor transactions by attention
- Clustering of transaction embeddings — groups transactions the model sees as similar
- UMAP — 2D visualization of transaction and customer embeddings
Google or a one-time email code. No password.