Information Reliability Service - a novel approach to fraud detection
summary
Swiss Post builds Iris, an Iris trust score service that flags risky addresses to stop account activation and cut ~5M CHF/year losses. It trains a Linear UCB contextual MAB on shipment records plus expert fraud labels: 72 features, a 3‑month 40GB snapshot with ~12M non‑fraud and ~10k frauds, balanced via under/oversampling and SMOTE. Data plumbing uses AWS Glue → S3 → Athena → SageMaker; reward engineering sets TP=+100, FN≈−60 to favor recall (achieves ~96% recall), and deployment is CI/CD–ready for 2025 rollout.
AI-written summary · about
chapters (12)
- 0:00Information Reliability Service
- 1:26Swiss Post fraud threats
- 3:57Project Iris trust score
- 6:52Multi-armed bandit model
- 10:48Data collection challenges
- 14:21Handling imbalanced datasets
- 19:18Reward engineering
- 23:32Experiment results
- 27:07Confusion matrix trade-offs
- 29:37Practical application
- 30:00Key takeaways
- 31:23Q&A
related talks
37mThe story of IRES - How Swiss Post fights fraud with ML on AWS
43mFrom Shadows to Spotlight - How Swiss Post Performs Reliable ML Deployment
39mFraud Detection: Using ML to Identify and Manage Fraudulent Activities
34mFrom Shadows to Spotlight - How Swiss Post Performs Reliable ML Deployment
42mBuilding a Robust Fraud Detection System
08mAI-Powered Fraud Detection in Financial System
40mFraud detection without labeled data
16mLeveraging ML to Create a Reliable Anti-Fraud System