Risk Entity Watch - Using Anomaly Detection to Fight Fraud

Becky Hui, Christopher Settles, Sahana Noru, Sergey Zelvenskiy28 Sep 2023

Problem

Uber runs supervised fraud models for dozens of use cases, each with its own label type: chargebacks, refund abuse, GPS spoofing. This becomes unmanageable as new lines of business add fraud classes that have no labels yet. The Risk team needs to flag suspicious entities without a label per fraud type. Human agents review every flag, and a wrong decision can deny a user service or income, so each flag needs a reason the agent can check.

Example from the article: a single trip request involves at least three entities: a rider, a driver and a payment method. The event changes their features, for example the trip count of each.

Solution in brief

Risk Entity Watch generates one grid of windowed metrics for every entity in every event type and trains an unsupervised anomaly model on it. HAIFA, a histogram check, then lists the features in which a flagged entity is isolated. Binary search sets its threshold so every flag gets at least one reason.

Questions this case answers
  • How to detect abusive entities with unsupervised anomaly detection when labeled examples are missing for each new abuse type
  • How to explain each individual anomaly flag to a human reviewer with an explainer cheap enough for streaming inference
  • How to flag fraudulent accounts, devices or payment instruments when new fraud types appear faster than labels?
  • How to build entity-level features for fraud detection without redesigning them for every entity and event?

Tools and methods

  • Entity Feature Generation (EFG) — computes each metric for every entity, window and event type
  • Unsupervised anomaly detection — scores entities against the normal population
  • HAIFA (Histogram Analysis of Important Features for Anomalies) — explains each flag, also used to select features and find noise
  • Bucket-threshold autotuning — binary search for the smallest threshold that explains all anomalies
  • Apache Hive — holds the training set
Continue reading — it's free

Google or a one-time email code. No password.