[KongchangAI]
ConceptTraining-Serving Skew / 训练服务偏差

Training/Serving Skew

A classic problem in machine learning engineering in which the feature distribution seen by a model during offline training differs from the feature distribution of actual inputs during online inference, arising from separate implementations of feature logic and differences in data arrival timing.

Timeline (last 90 days)

Nov 6

In fraud detection, training/serving skew is especially fatal, because fraud patterns themselves evolve rapidly, and a small drift in feature definitions may directly render the model ineffective.

Unverified50%
Nov 6

Common industry measures to mitigate training/serving skew include a single source of truth for feature definitions, offline/online feature consistency checks, and feature monitoring and alerting.

Unverified50%

All Facts (2)

Source Articles