Creating & operating ML models from event-based data using feature stor...
summary
Claims a practical stack: a time-travel feature engine (Kaskada) that loads raw event streams like parquet files in S3 and runs in Jupyter to compute features directly from event-based data. Shows how to express event-relative feature definitions, join across entities at precise event times without leakage, and materialize discoverable features into feature stores or caches. Reports game metrics (day‑one retention ≈25%, ≈94% churn in month one) and says iteration drops from months to days/weeks for backfills, experiments, and production validation.
AI-written summary · about
chapters (13)
- 0:00Intro
- 0:19A true story
- 1:22Getting the impact
- 2:19Timebased feature engine
- 3:49Machine learning process
- 5:06Building features
- 6:21Kaskada
- 7:06What you need
- 8:15Data Science Experience
- 9:00Feature Definitions
- 10:35Instant iteration
- 12:16Joins
- 15:35Summary
related talks
40mCreating and operating ML models from event based data using feature stores
40mOvercoming data infrastructure limitations with a new paradigm for ML
30mHow Feature Stores Enable Operational ML
45mBuilding a Feature Platform to Scale Machine Learning
01h 28mFeature Engineering Made Simple
01h 52mDeploying a Fraud Detection Model with a Feature Store
10mFast Sub-ML Usecase Development using Feature Stores
15mFeature Stores: Powering Machine Learning Infrastructure with Apache Flink