No labels? No problem!
summary
A production ML pipeline that labels fraud without true labels uses a two-step proxy labels system: first find suspect intervals with temporal anomaly detection, then extract precise attack clusters with graph-based pattern detection linking transactions by shared identifiers. The deployed model scores hundreds of transactions per second (millions per hour), verifies proxy labels against a domain-expert ground truth set, reports roughly 100% precision and 97% recall on that testbed, and keeps improving via a human feedback loop.
AI-written summary · about
chapters (11)
- 0:00Ajen and fraud detection intro
- 2:15Global credit card fraud rise
- 7:10Static rules vs fraudsters
- 8:12Missing labels in fraud detection
- 10:42Proxy labeling system design
- 14:13Temporal anomaly detection
- 16:19Graph-based transaction patterns
- 21:10Evaluating proxy labels
- 22:54Handling false negatives
- 24:18Key takeaways on proxy labeling
- 26:48Q&A
related talks
33mWeak supervision: a new paradigm for unreliable data
40mFraud detection without labeled data
16mLeveraging ML to Create a Reliable Anti-Fraud System
36mContinuous Monitoring of Machine Learning Models
25mReal-Time Fraud Detection at Scale
12mCloud-Native Friendly Fraud Detection
52mModern Fraud Prevention Using Deep Learning
30mData Driven Security