A novel weakly supervised problem: Learning from positive-unlabeled proportions

Abstract

Standard supervised classification learns a classifier from a set of labeled examples. Alternatively, in the field of weakly supervised classification different frameworks have been presented where the training data cannot be certainly labeled. In this paper, the novel problem of learning from positive-unlabeled proportions is presented. The provided examples are unlabeled and the only class information available consists of the proportions of positive and unlabeled examples in different subsets of the training dataset. An expectation-maximization method that learns Bayesian network classifiers from this kind of data is proposed. A set of experiments has been designed with the objective of shedding light on the capability of learning from this kind of data throughout different scenarios of increasing complexity.