Label Shift

Label shift occurs when class proportions change between training and deployment data while the input distribution within each class is assumed to remain stable. Common approaches estimate target class priors from unlabeled predictions or importance weights, then adjust posteriors, train with reweighting, or adapt models at test time. The methods are used in domain adaptation, online and zero-shot learning, and conformal prediction, with evaluations on image, text, medical, and scientific data. A central challenge is that prior estimates depend on classifier calibration and can be unreliable when class-conditional distributions also change; newer work extends the theory to continuous labels and studies uncertainty and validity guarantees.