Deploying a Fraud Detection Model with a Feature Store
summary
Builds a real-time fraud pipeline end to end using a Tecton feature store on synthetic PSIM transactions: defines features (credit_score, age_in_days, days_since_last_tx), implements sliding-window aggregations (counts and mean amounts over 1h/12h/3d/7d), backfills historical values with Spark jobs, and exposes a feature service that materializes online values. Trains a simple model locally with XGBoost + MLflow -> SageMaker endpoint. Online storage uses DynamoDB; offline uses Delta on S3. Live demo shows materialization latency hiccups but the pipeline and serving pattern work.
AI-written summary · about
chapters (9)
- 0:00Introduction to workshop
- 10:57Fraud detection overview
- 21:47Building first features
- 37:04Aggregation features
- 50:25Feature service setup
- 1:00:19Training model
- 1:12:53Adding new features
- 1:28:56Troubleshooting demo
- 1:32:06Q&A session
related talks
14mLearn Vertex AI while building a fraud detection system
01h 12mBuild Simpler Production ML Systems using Feature/Training/Inference Pipelines
30mAmazon SageMaker Feature Store: Store, discover, & share features for ML apps
15mFeature Stores: Powering Machine Learning Infrastructure with Apache Flink
25mReal-Time Fraud Detection at Scale
20mNavigating the Skies of Data: A Flight Through Feature Stores
18mDesigning a scalable ML Ops Pipeline: Insights and best practices
32mA feature Store for Machine Learning