Problem
Booking.com fights payment fraud, fake hotels and reviews, and marketing abuse. Fraudsters change tactics often, so the system must use new information immediately. Fraud attacks are usually connected: identifiers such as an email, card or account appear in several requests. A model that scores each request alone cannot see that a clean-looking request is linked to earlier fraud. Scoring is synchronous, so any extra lookup must be fast.
Example from the article: a reservation request arrives from an account and a card that were never flagged, but a few links away another card was reported as fraudulent by its issuing bank.
Solution in brief
Store every observed identifier as a graph node and connect identifiers seen together in one request. At scoring time, build a small local graph around the request's identifiers and turn it into numeric features such as node counts and the distance to a fraud notice. A model or an expert rule then maps those features to a risk estimate.
Questions this case answers
- How to detect coordinated abuse when each single request looks clean but its identifiers are linked to known bad actors
- How to compute graph-structure features over a large history of identifiers within the latency limit of a synchronous decision
- How to use links between accounts, cards and other identifiers to score the fraud risk of a reservation request?
- How to turn a graph of linked identifiers into features a fraud model or a rule can consume?
Tools and methods
- Identifier graph — nodes are identifiers (account, card), edges link identifiers seen together, fraud notices are extra nodes
- Graph features — counts of nodes by type and the shortest distance to a fraud notice
- Breadth-first search — fetches the connected network around the request identifiers
- JanusGraph with a Cassandra backend — graph storage and queries
- Graph Service — builds the features for the Fraud Detection Service
Google or a one-time email code. No password.