Problem
An anti-fraud team in a bank relies on hand-built if-then rules. Fraudsters adapt, so old rules start hitting good customers and every change means a rework. Each alert is checked manually: 40 rules with 15 alerts a day each already give 600 cases, which burns out staff. False blocks lower customer trust, and fraud is under 1% of transactions, so labels are scarce and noisy.
Example from the article: a 100 thousand ruble transfer through the Faster Payments System (SBP) is atypical for an elderly customer who rarely transfers money, but normal for a young sole proprietor who often makes large SBP transfers.
Solution in brief
Score how well a transaction fits the customer's own behaviour. The speaker compares two models: a CatBoost classifier trained on rule alerts, and a model that predicts the type of the next transaction from the last N transactions. The second one needs no labels and was chosen. Both sit after the rules in a cascade.
Questions this case answers
- How to flag atypical events in an entity's action sequence by modelling its own history instead of fixed rules when the fraud label is rare and noisy
- How to predict the type of an entity's next event from its last N events and treat a low probability as a deviation, without labels
- How to flag atypical customer transactions by modelling each client's own behaviour instead of fixed rules
- How to cut false positives of anti-fraud rules with a model score placed after the rules
Tools and methods
- CatBoost — classifier on rule alerts (approach 1)
- Next-transaction-type prediction — multiclass model over a sliding window of N transactions (approach 2)
- Hit rate and NDCG — offline metrics for approach 2
- Rule cascade — model score as a filter on rule alerts or as an extra rule condition
Google or a one-time email code. No password.
