Creating and operating ML models from event based data using feature stores
summary
Claims that a feature engine with native time travel (Kaskada) lets teams compute historical feature values instantly, test different prediction and label times, and run temporal joins across entities so models from event streams iterate in days or weeks instead of the typical 7–12 months. Demoed components: backfilled historical feature generation, expressive time-selection APIs, and a feature store + low-latency cache (Redis) for serving. Result: reproducible training examples without leakage and faster productionization of features; built and used in production scenarios, shortening iteration and fixing join/labeling errors.
AI-written summary · about
chapters (8)
- 0:00Introduction to talk
- 1:27Story of retention problem
- 8:28Event-based data model goals
- 12:49Iterative feature engineering challenges
- 22:18Time travel in feature engineering
- 25:47Getting features to production
- 34:20Key takeaways and solutions
- 37:29Q&A on applications
related talks
17mCreating & operating ML models from event-based data using feature stor...
21mTime for quality ML from event-based data
30mHow Feature Stores Enable Operational ML
19mML feature store a cornerstone in modern data architecture for AI/ML
15mFeature Stores: Powering Machine Learning Infrastructure with Apache Flink
32mA feature Store for Machine Learning
08mFeature Store Usage Patterns - From a Single Data Scientist to an Enterprise
20mSemblance: Feature Generation In Real-Time and Batch Without Time-Travel