Fraud detection without labeled data
summary
Claims that unsupervised anomaly detection plus streaming stats catches fraud without labeled data, replacing costly manual review. Tools and patterns: sliding-window HyperLogLog and count-min sketch to shrink state, Kafka-backed rules engine and Redis/Smite DB for low-latency updates. Concrete numbers: businesses report ~10x more loss from false positives than from fraud, and simple rules like “>10 unique cards per IP” become a percentile-based top-1% alert that adapts in real time. This stack runs in production and reduces reviews while keeping latency low.
AI-written summary · about
chapters (28)
- 0:00Introduction
- 2:29Overview
- 6:34Examples of fraud
- 10:06Manual review
- 13:24Automatic classification
- 15:50Automated rules
- 18:00Machine learning
- 18:21Machine learning hammer
- 20:52The worst part
- 21:04Label data
- 22:09adversarial problem
- 22:52deep learning
- 23:07simple rules
- 23:22example
- 24:28advantages
- 24:56spam detection
- 26:46data warehouse
- 28:16infrastructure overview
- 28:35challenges
- 29:04hacksoul
- 30:09storage size
- 31:35compression
- 33:01disk latency
- 34:12high availability
- 35:41sharding
- 36:22Smite DB
- 36:48Perceptual hashes
- 38:31Recap
related talks
13mHarnessing Machine Learning to Combat E-commerce Fraud
13mBuilding Financial Fortresses: Real-Time Fraud Detection
51mReal-time Fraud Detection: Challenges and Solutions
10mReal-Time Fraud Prevention at Scale
16mLeveraging ML to Create a Reliable Anti-Fraud System
18mReal-time Fraud Detection using Deep Learning
13mBehavioral Analytics for Real-Time Fraud Prevention
39mFraud Detection: Using ML to Identify and Manage Fraudulent Activities