Importance Weighting Adapts Unlabeled-Unlabeled Learning to Distribution Shift

This method applies importance weighting to Unlabeled-unlabeled (UU) learning to mitigate test-time distribution shifts. It enables binary classifier training from two unlabeled datasets with different class-priors by minimizing test risk via estimated importance weights.