Deep-PUMR: Deep positive and unlabeled learning with manifold regularization

Abstract

Training a binary classifier only on positive and unlabeled examples (i.e., the PU learning) is an important yet challenging issue, widely seen in many problems in which it is difficult to obtain negative examples. Existing methods for handling this challenge often perform unsatisfactorily, since they often ignore the relations between positive and unlabeled examples and are also limited to the traditional shallow learning frameworks. Therefore, this work proposes a new approach: Deep Positive and Unlabeled learning with Manifold Regularization (Deep-PUMR), which integrates the manifold regularization with deep neural networks to address the above issues with classic PU learning. Deep-PUMR holds two major advantages: (i) Our method exploits the manifold properties of data distribution to capture the relationship of positive and unlabeled examples; (ii) The adopted deep network enables Deep-PUMR with strong learning ability, especially on large-scale datasets. Extensive experiments on five diverse datasets demonstrate that Deep-PUMR achieves the state-of-the-art performance in comparison with classic PU learning algorithms and risk estimators.