When is Black Box Shift Estimation identifiable?

Question: Black Box Shift Estimation is used to infer target-domain class proportions from unlabelled target data. Which combination of assumptions is central to the original method? A) The conditional label distribution P(y|x) is invariant, and the classifier is perfectly calibrated B) The covariate marginal P(x) is invariant, and the classifier has zero training error C) Both P(x|y) and P(y) remain invariant, while P(y|x) changes D) P(x|y) is invariant across domains, and the black-box predictor's confusion matrix is invertible Correct: D Explanation: BBSE addresses label shift, under which class priors change while the class-conditional feature distributions remain stable. Its moment system can identify the new class proportions when the relevant confusion matrix is invertible. Perfect calibration and perfect accuracy are not required. Topic: advanced ML / distribution shift / label shift