Project RADAR: Intelligent Early Fraud Detection System with Humans in the Loop

Tiffany Yu, Edwin Ng, Sergey Zelvenskiy, Garvit Harisinghani, Robin Wei1 Feb 2022

Problem

Uber pays its partners even when a rider's payment fails, so the marketplace carries the payment fraud risk. Two loss types matter: "do not settle" (DNS), where the payment for a completed trip could not be collected, and chargebacks, where the user disputes a collected payment. Losses for one order day keep changing for weeks, so an early spike may be fraud or just more activity. Analysts write blocking rules by hand, which is slow, and supervised models miss new attack patterns.

Example from the article: a jump in immature losses in one marketplace segment, where RADAR must judge whether it will stay serious at full maturity.

Solution in brief

RADAR Detect forecasts each segment's loss time series and flags deviations as attacks. It projects the flagged losses to full maturity and sends only the severe ones to analysts. RADAR Protect mines frequent feature combinations among fraudulent orders with FP-Growth and validates them as precise rules. An analyst approves each rule before it goes live.

Questions this case answers
  • How to detect the start of an adversarial attack from delayed, still-changing loss data and turn it into reviewable blocking rules
  • How to prioritize anomalies by their projected final loss so that reviewers are not flooded with alerts
  • How to spot the start of a payment fraud attack when the losses are still immature and keep changing?
  • How to decide which anomalous loss spikes deserve an analyst's attention?

Tools and methods

  • Orbit KTR (Kernel-based Time-varying Regression) — decomposes each time series into trend, seasonality and regression parts to set the anomaly baseline
  • Smart thresholds — forecast of final losses, used to rank anomalies
  • FP-Growth — mines common feature-value sets among fraudulent orders to produce rule candidates
  • Customized MRMR — automatic feature selection before pattern mining
  • Apache Spark, Kafka, Hive, Jira — pipelines and analyst task queue
Continue reading — it's free

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