Model drift
A decline in a model’s predictive quality as data or relationships change during use.
Why can an unchanged model become less suitable?
Model drift refers to declining predictive quality as data or relationships change during use. The stored model weights do not have to change for this to happen. Concept drift more specifically describes a change in the relationship between inputs and the target to be predicted; data drift concerns the distribution of inputs. These distinctions help investigate causes. A changed data mix alone does not establish a loss of predictive quality.
In a fictional service business, a new maintenance procedure shortens actual work duration for the same previously recorded job features. An older forecasting model could therefore increasingly predict visits that are too long. Reviewers compare current predictions with checked actual durations, including the affected job types. Before making an adjustment, they also investigate incorrect time units, incomplete work reports and changed recording rules. If a new model variant is trained, it is assessed using current held-out examples and separately approved before adoption. The response depends on the cause identified.
Sources
Related terms
- Machine learningModels learn patterns from data and apply them to new inputs.
- Forecasting modelA model that uses available information to estimate a future quantity for a defined period.
- Training dataThe examples from which a model learns relationships or parameters during training.
- Test data in machine learningData reserved for final assessment whose results were not used to select the model variant being assessed.
- Explainable AIApproaches that make an AI system’s results or operation understandable to the intended recipients.
- Data driftA change in the distribution of model inputs relative to an explicitly chosen reference.
Your project with ASES
Describe your project to us. We will get back to you with an initial assessment.