Concept Drift
The change in the relationship between input features and the target variable over time.
Overview
Concept drift occurs when the statistical relationship between model inputs and the target output changes over time, causing a model's predictions to become less accurate even though the input distribution may remain similar. This differs from data drift, where the input distribution itself changes.
Key Details
For example, consumer purchasing behavior shifted dramatically during COVID-19 — the relationship between features (demographics, browsing history) and purchases changed, even though the feature distributions may not have. Detecting concept drift requires monitoring prediction accuracy on labeled production data, which is often delayed. Strategies include periodic retraining, online learning, and ensemble methods that weight recent data more heavily.