Concept Drift
A shift in the relationship between input features and the target variable, causing a model to become less accurate even if input distributions are unchanged.
Concept drift (also called posterior drift) is the most insidious form of model degradation: the underlying relationship that the model learned during training changes in the real world, rendering the model's parameters incorrect — even if the inputs look statistically similar to training data.
Example: A fraud detection model learns that transactions from a specific merchant category are low-risk. A new fraud ring specifically targets that merchant category, exploiting the model's blind spot. The input distribution of merchant categories may not change significantly (data drift is low), but the relationship between "this merchant category" and "fraud probability" has fundamentally changed.
Types of concept drift:
- Sudden drift — Abrupt change in the data-generating process (e.g., COVID-19's effect on spending models)
- Gradual drift — Slow, continuous change over months or years (e.g., inflation eroding the predictive power of absolute income thresholds)
- Recurring drift — Cyclic changes (e.g., seasonal patterns that the model did not see in training)
- Incremental drift — Steady, linear shift in the concept (e.g., loan default rates slowly rising with interest rates)
Detection challenge: Concept drift cannot be detected from inputs alone — it requires ground-truth labels. For time-lagged outcomes (30-day delinquency, 12-month default), there is an inherent detection delay. Explanation-based monitoring (watching for feature importance shifts) and distribution monitoring (PSI) provide earlier signals but are leading indicators, not direct measures.