Conceptfeature drift / data drift
Feature Distribution Drift
The phenomenon in which the statistical distribution of input features during online inference deviates from the distribution at training time, causing degradation in model performance; it is addressed through monitoring, alerting and retraining.
Timeline (last 90 days)
Oct 23
Coping with feature drift typically requires continuously monitoring feature distributions, setting alert thresholds, retraining periodically or on trigger, and keeping version snapshots of training data and feature definitions.
Unverified50%