Glossary · IIoT, data and AI
Concept drift
German: Konzeptdrift
In machine learning, concept drift is a change over time in the relationship between a model's input data and the target it predicts, so that a mapping learned from past data no longer holds. The inputs may look unchanged while their meaning for the outcome has shifted.
- IIoT
- AI
In one sentence
Concept drift is a change in the relationship between inputs and the predicted target, making a trained model's learned mapping outdated.
Example
After a supplier changes the resin formulation, the same temperature and pressure profile no longer predicts part quality, although the sensor values look familiar.
How it applies
- Engineering: Concept drift can be caused by material changes, tool wear, new product variants, process modifications or changed quality criteria. Record such changes so they can be linked to model behavior.
- Operation: Concept drift is detected by comparing predictions with actual outcomes, which requires ground truth such as lab results or inspection data. Input monitoring alone does not reveal it.
- Documentation: State the conditions under which the model was trained and validated, and list process changes that require revalidation or retraining (Model validation).
Concept drift vs. data drift
Data drift is a change in the distribution of the inputs; the learned relationship may still be valid. Concept drift is a change in the relationship itself; the inputs may be unchanged. Both can lead to Model drift, the observable decline in model performance, and both can occur at the same time.