PU Learning using Optimal Transport with Laplacian Regularization

Abstract

PU learning is one of the fields of machine learning and is an extension of binary classification. It differs from binary classification in that only positive-labeled and unlabeled data is given as training data. In PU learning, there is an assumption that similar datas have close probability of belonging to a positive class. One of the methods of PU learning is to use the partial optimal transport (POT) problem, but this method does not take into account this assumption. To this end, this paper proposed the POT with Laplacian regularization to perform mapping based on the distance relation before and after transportation. Numerical evaluations show the effectiveness of our proposed method.