Problem
Grab keeps adding products: ride hailing, food and grocery delivery, financial services. Each new product brings new fraud patterns, and a new model for each one is too slow. A model built per pattern is slow to ship and leaves the new product unprotected at the start (cold start). Rules cover only simple fraud. Decision-tree models need good labels and hand-built features, and fraud labels are scarce and biased.
Example from the article: a fraud syndicate shares physical devices, or a merchant colludes with an isolated set of passengers through food orders.
Solution in brief
Build one graph from the entities and their shared properties, then train a Relational Graph Convolutional Network (RGCN) on it in a semi-supervised way. Only a few percent of nodes carry labels, and the model spreads information along the edges to score every node. Dense clusters of accounts sharing devices show up for the model and for an investigator.
Questions this case answers
- How to detect coordinated abuse when labels cover only a small share of entities and new abuse patterns keep appearing
- How to score entities for risk using the relationships between them instead of hand-built per-entity features
- How to detect fraud when only a few percent of accounts are labeled?
- How to catch new fraud patterns without building a new model for every new product?
Tools and methods
- Entity graph — nodes are passengers, merchants, devices and similar entities; edges are shared devices, Wi-Fi routers, identities, delivery addresses and orders.
- Relational Graph Convolutional Network (RGCN) — semi-supervised node classifier that outputs a fraud probability for each node.
- Node features from domain knowledge — account attributes such as age and number of orders.
- Graph visualiser — shows dense device-sharing clusters to explain a score.
Google or a one-time email code. No password.