Building a Robust Fraud Detection System
summary
Streaming clickstream and profile data feed three classifiers — dynamic labeling, behavior-similarity and email-similarity — whose outputs are OR’ed to flag fraud on DISQO’s Service Junkie panel. Cosine similarity on n-grams and Levenshtein distance on local-part emails detect coordinated accounts; target encoding and bucketing turn session-level clicks into user features. A convolutional neural network yields the best detection (recall jumps ~0.05→0.51), XGBoost recall ~0.17→0.41, training weekly on SageMaker (≈6 hours, <$20/month), daily batch scoring (~30 min preprocessing, <$10/month).
AI-written summary · about
chapters (23)
- 0:00Introduction
- 1:24Welcome
- 2:15Disco
- 2:43Service Junkie
- 3:04Agenda
- 3:19Why is it rising
- 4:52Types of Fraud
- 5:23Robust Flow Detection System
- 5:49Types of Data
- 6:23Models
- 6:35Dynamic Labeling
- 9:50Feature Engineering
- 13:07Target encoding
- 14:53Feature selection
- 15:36Evaluation Metrics
- 19:46Behavior Similarity
- 25:01Email Similarity
- 26:58Combining Predictions
- 27:13Deployment
- 28:39Prediction
- 29:51Summary
- 30:29Questions
- 37:31False Positive vs False Negative
related talks
13mBehavioral Analytics for Real-Time Fraud Prevention
16mAI-Driven Behavioral Analytics for Real-Time Fraud Prevention
16mLeveraging ML to Create a Reliable Anti-Fraud System
39mFraud Detection: Using ML to Identify and Manage Fraudulent Activities
25mReal-Time Fraud Detection at Scale
18mReal-time Fraud Detection using Deep Learning
13mHarnessing Machine Learning to Combat E-commerce Fraud
47mTrain, Evaluate, Repeat: Building a Credit Card Fraud Detection System