Key takeaways
- The core feature store concepts: feature sets, offline versus online stores, feature versioning, and derived parent and child feature sets
- What's new in 1.0, including near UI and CLI parity, Snowflake as an alpha offline store option, key pair authentication, and a 100x online store throughput improvement
- How AT&T uses the feature store in production, from a self-service rollout with tens of thousands of onboarded features to 50-plus fraud models and a 34% model lift
- How AT&T's churn model team found a feature set that lifted model performance by 11%, and had it in production within about a week of that discovery
- Built-in enterprise governance: role-based access control, an internal review process for publishing feature sets, and anonymization for sensitive PCI, PII, and SPI data
- A live demo registering a feature set, ingesting data via UI and notebook, and retrieving features through downloads, Spark, and the online store
Overview
This recording covers the general availability of H2O Feature Store 1.0, with H2O's Chris Simmatt and Kuba joined by AT&T's Prativa Sugerman. It covers core concepts, feature sets, offline and online stores, versioning, and derived feature sets, plus what's new in 1.0: near UI and CLI parity, an alpha Snowflake offline store, key pair authentication, and a 100x online store throughput gain. Sugerman shares how AT&T built tens of thousands of features into the store, cutting fraud losses with 50-plus production models and lifting a churn model within about a week.