Skip to content
ML Academy

Introduced on day 13

Prior correction

Rescaling a classifier's odds by the ratio of the deployment class prior to the training prior, so probabilities match the base rate in use.

In detail

After class weighting or resampling to a balanced prior, predicted odds are inflated by the factor (1 - pi) / pi, with pi the real prevalence. Multiplying the odds by pi / (1 - pi) undoes the weighting, and the same formula adapts a model to a changed prevalence under label shift. It leaves the ranking, and therefore ROC-AUC, unchanged.