Label Shift Explainer
Label shift occurs when the proportions of target classes differ between training and deployment. The standard assumption is that the distribution of inputs within each class remains stable even though the class frequencies change.
A screening model may be trained where positive cases are common but deployed where they are rare, or the reverse. This can change predictive values and decision thresholds even if class-conditional patterns remain similar.
Because production labels may be delayed, methods estimate new class proportions from model predictions and a validation confusion matrix. Estimates depend on the model separating classes well enough and on the label-shift assumption being approximately true.
Teams can adjust predicted probabilities or thresholds, then validate performance with real target labels when available. If features within a class also changed, a label-shift correction alone may be misleading and a broader investigation is needed.
Label shift occurs when the proportions of target classes change between training and deployment while class-conditional input patterns are assumed to remain stable.
Key terms include label shift, class prior, target proportion, class-conditional distribution, confusion matrix, probability adjustment and threshold.
The target population contains a different balance of outcomes than the training set.
That changed base rate can alter how model scores should be interpreted.
Estimators assume inputs within each class remain sufficiently stable.
Target labels should be checked when available to confirm that adjustment improved the real task.
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.