PU Learning
A branch of weakly supervised learning that handles settings where only positive examples are labeled in the training set, while the remaining data may be either positive or negative. Classical solutions include class-prior estimation with reweighting and training with unbiased risk estimators.
Timeline (last 90 days)
PU Learning cannot simply treat unlabeled samples as the negative class, otherwise it introduces severe label noise.
Classic PU Learning solutions include: estimating the class prior and then reweighting unlabeled samples, and treating unlabeled data as a mixture of positive and negative distributions and training a classifier via an unbiased risk estimator.
All Facts (2)
PU Learning cannot simply treat unlabeled samples as the negative class, otherwise it introduces severe label noise.
50%UnverifiedClassic PU Learning solutions include: estimating the class prior and then reweighting unlabeled samples, and treating unlabeled data as a mixture of positive and negative distributions and training a classifier via an unbiased risk estimator.
50%