Definition
Concept drift is a change over time in the relationship between inputs and the correct target, label, or action. The same observed features no longer support the same decision because the underlying process, behavior, policy, or environment has changed.
Drift can be sudden, gradual, recurring, or limited to part of the population. Detecting it usually requires fresh outcome labels or another source of ground truth; monitoring inputs alone cannot establish that the correct relationship changed.
Response depends on the cause. Retraining may help when behavior has moved, but it can entrench a bad policy change or temporary anomaly. Investigate the affected slice and the real-world process before updating the model.
Distinguish it from nearby terms
Data drift changes the distribution of inputs. Concept drift changes what those inputs mean for the target. Model drift describes changes in model behavior or performance and may result from either one.
Check your understanding
A fraud rule learned that a particular purchase pattern was suspicious. Customers adopt the pattern after a new payment product launches, so the same features no longer imply fraud. That is concept drift.