Learning from Candidate Labeling Sets