Tag
The paper proposes an importance weighting method for unlabeled-unlabeled learning to adapt to distribution shifts, applicable to various weak supervision problems without assumptions on shift types.