Covariate Shift Explainer
Covariate shift is a specific kind of dataset shift. The distribution of input features changes between training and use, while the conditional relationship between those inputs and the target is assumed to stay the same.
For example, a model trained on one mix of ages or devices may later receive a different mix. Even if the prediction rule is unchanged, performance can move because the model now encounters some regions more often.
Methods for addressing covariate shift often estimate how common each input region is in the source and target data, then reweight examples. This works only when the assumption is reasonable and important target regions have training support.
Real deployments can contain several shifts at once, so covariate shift should not be declared from a feature histogram alone. Teams test assumptions, inspect labels when available, and watch for new inputs outside the training domain.
Covariate shift is a distribution change in model inputs while the conditional relationship between inputs and the target is assumed to remain stable.
Key terms include covariate shift, input distribution, source domain, target domain, conditional relationship, density ratio, weighting and support.
Some feature combinations become more or less common in the target population.
The standard assumption keeps the target relationship conditional on those features unchanged.
Training examples must cover the regions that matter in the target domain.
Weights cannot create reliable evidence for inputs the training data never represented.
Explore more "Explainers"
Discover additional explainers across politics, science, business, technology, and other fields. Each explainer breaks down a complex idea into clear, everyday language—helping you better understand how major concepts, systems, and debates shape the world around us.