Train, Evaluate, Repeat: Building a Credit Card Fraud Detection System
summary
Describes concrete solutions for Stripe’s fraud stack: chooses merchant‑agnostic labels like chargebacks and issuer fraud records, experiments to set a 10% target rate for positives instead of fixed downsampling, and implements a uniform holdback group (random X%) to get unbiased labels. Uses gradient‑boosted trees and scikit‑learn sample weights, handles <1% fraud prevalence, retrains frequently, and shows the holdback reduces estimator variance and preserves long‑term performance.
AI-written summary · about
chapters (8)
- 0:00Stripe and credit card fraud
- 3:26Defining fraud impact
- 7:00Labeling fraud data challenges
- 12:05Class imbalance handling
- 19:38Training data target rates
- 23:17Counterfactual evaluation
- 35:46Key takeaways and optimization
- 37:10Q&A
related talks
39mTrain, Evaluate, Repeat: Building a Credit Card Fraud Detection System
13mHarnessing Machine Learning to Combat E-commerce Fraud
36mData Quality Assurance in Credit Card Fraud Detection
40mFraud detection without labeled data
45mAnti-Fraud in Reverse and the Use of ML Methods in It
42mBuilding a Robust Fraud Detection System
26mAI-Powered Fraud Detection in Cloud-Native Digital Payments
33mChris Stucchio of Simpl at Machine Learning Developers Summit 2019