Glossary · IIoT, data and AI
Model drift
Also known as: Model decay, Model degradation
German: Modelldrift
In machine learning, model drift is the decline of a deployed model's predictive performance over time because the data or conditions it encounters in operation differ from those it was trained on. It is usually caused by data drift, concept drift or both.
- IIoT
- AI
In one sentence
Model drift is the decline of a deployed model's performance over time as operating data and conditions move away from its training data.
Example
A model predicting energy consumption becomes steadily less accurate after a new product mix is introduced, and its error doubles within three months.
How it applies
- Engineering: Plan for drift from the start: define performance metrics, monitoring, thresholds and a retraining process before the model goes live.
- Operation: Drift is detected through Model monitoring: tracking prediction error where ground truth is available, and input distributions where it is not.
- Documentation: Document the validated operating range and the conditions that trigger revalidation. Treat a retrained model as a new version under Change control.
Model drift vs. sensor drift
Sensor drift is a gradual change in a sensor's output for the same physical value, a measurement problem. It can cause model drift, because the model receives shifted inputs. Distinguishing the two avoids retraining a model on faulty measurements instead of recalibrating the sensor.