What is concept drift?
Concept drift is a change in the relationship between a model’s inputs and the outcome it predicts. A model learns a mapping, such as from a home’s features to its sale price. When that mapping changes in the world, the model keeps applying the old one, and its predictions go wrong even though the inputs look as they always did.
Concept drift and data drift are the two main causes of model drift. In data drift the mix of inputs changes while the relationship holds. Concept drift is harder to catch because input statistics do not reveal it. It shows up only when predictions are compared with actual outcomes, which may arrive weeks or months later. It can be sudden, as when a new regulation changes buyer behavior, or it can build gradually or recur with the seasons.
Concept drift in practice
A county assessor’s office uses a model to estimate the market value of homes from their features and recent sales. When mortgage rates climb, buyers can afford less, and homes start to sell below the model’s estimates. Square footage, location, and condition are unchanged, so input monitoring raises no alarm. The office catches the gap by comparing each month’s sale prices with the model’s estimates, then retrains with recent sales weighted more heavily.
Fraud detection shows the problem in a sharper form. Once a model blocks one scheme, fraudsters move to another, so the patterns that signal fraud keep changing and the model needs regular retraining on confirmed cases.