Cloud-Native Friendly Fraud Detection
summary
Combines temporal graph attention networks with a transformer sequence model to flag friendly fraud, reporting 67–78% detection versus 38–47% for traditional ML and 23–31% for rule systems, and a 35–45% reduction in false positives. Operates on payment graphs of 50 to 100 million accounts, 15 to 25 million merchants and 10 to 50 billion edges with 100–500 million new connections per day. Uses distributed graph databases, segment-level calibration, Kubernetes autoscaling and human-in-the-loop feedback to keep latency low and models current.
AI-written summary · about
chapters (13)
- 0:00Intro and Goal
- 0:58What Is Friendly Fraud
- 1:24Why Detection Is Hard
- 2:14Limits of Current Methods
- 3:06Hybrid AI Framework
- 3:53Payment Graph at Scale
- 4:41Graph Structure and Attention
- 6:12Transformer Behavior Modeling
- 6:39Metrics and Benchmarks
- 8:14Deployment at Issuer Scale
- 8:59Cloud Native Architecture
- 9:55Human in the Loop Feedback
- 10:45Business Impact and Wrap Up
related talks
14mCloud-Native AI for Fraud Prevention with Bayesian Networks
26mAdvancing Financial Fraud Detection with Graph Dbs
25mReal-Time Fraud Detection at Scale
26mAI-Powered Fraud Detection in Cloud-Native Digital Payments
31mLet's do graph machine learning on Cloud
28mA Quantum Leap In Fraud Detection Accuracy Using Graph-based Features
16mLeveraging ML to Create a Reliable Anti-Fraud System
52mModern Fraud Prevention Using Deep Learning