Data drift
A change in the distribution of model inputs relative to an explicitly chosen reference.
What has changed in the incoming data?
Data drift means that the distribution of model inputs changes relative to a chosen reference. Training data or an earlier period of use can provide the comparison; the reference used must be defined. Properties such as category frequencies or value ranges are examined. A drift signal points to a possible change. Whether it causes more incorrect predictions needs a separate assessment. A change in the relationship between inputs and the target cannot be inferred from the signal alone either.
In a fictional image-classification system, the share of poorly lit photos increases after a new capture device is introduced. The reference period and observed capture conditions are documented for the comparison. Reviewers then examine affected images and their checked class assignments to investigate possible changes in errors. They also examine whether capture practices actually changed or whether a transfer or processing error is involved. An unusual data comparison is a reason for that investigation; it proves neither a device defect nor the need for a new model.
Sources
Related terms
- Machine learningModels learn patterns from data and apply them to new inputs.
- Computer visionMethods for extracting information from images and video, such as classifying objects, locating them or distinguishing image regions.
- 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.
- AI biasSystematic distortions in AI systems that can disadvantage particular cases or groups of people.
- Model driftA decline in a model’s predictive quality as data or relationships change during use.
Your project with ASES
Describe your project to us. We will get back to you with an initial assessment.