Training-serving skew
the failure mode where a feature's value at training time differs from its value at prediction time in production, because the two paths compute it differently or draw from unsynchronized data, silently degrading a model's real-world accuracy below what its offline evaluation predicted.
Defined in 2 GAGE programs, which carry 3 distinct definitions of it. The wording above is taught in AI Data Governance: The Data Chair.
How each discipline defines it
The same term does different work depending on who is using it. These are the definitions as each program teaches them, unedited.
the failure mode where a feature's value at training time differs from its value at prediction time in production, because the two paths compute it differently or draw from unsynchronized data, silently degrading a model's real-world accuracy below what its offline evaluation predicted.
When a feature is computed differently in training than in production (different time windows, sources, or definitions), so the live model is effectively fed different data than it learned from. A leading, hard-to-diagnose cause of production degradation.
A silent failure in which a model performs well in testing but poorly in production because a feature was computed differently during training than at live inference. A feature store prevents it by serving the same feature definition to both.
Where it is taught
The exact lessons this term appears in. The first 7 topics of every program are free with a free account.
- The feature store: one governed source for every model, or chaos per project · Feeding the Machines, AI Data Governance: The Data Chair
- Enterprise Integration Strategy · Technology, Platforms and Vendors, Business AI Transformation
- Analytics, Telemetry and Observability · Technology, Platforms and Vendors, Business AI Transformation
Terms it appears with
Not an alphabetical neighbourhood: these are the terms taught in the same lessons, ranked by how often they appear together.