Tag
This paper proposes a generalized distribution-free semi-supervised learning framework that constructs unbiased risk estimators via linear combinations of component risks, extending PNU learning to multiclass classification while achieving lower variance and providing generalization bounds.