Asymmetric Loss for Positive-Unlabeled Learning

Abstract

Positive-unlabeled (PU) learning is a learning paradigm when only positive and unlabeled data are available in the training stage. This paradigm is particularly useful for the applications that the negative samples are hard to define or expensive to obtain. We propose a simple yet effective and scalable method to address PU learning with a novel asymmetric loss. The proposed asymmetric loss behaviors differently for the prediction errors of the labeled and unlabeled samples, and thus encourages the identification of the negative examples from the unlabeled set. For the PU learning with SCAR assumption, neither hyper-parameter nor class prior is required to be tuned or known. For the situation with selection bias on the labeled samples, we propose a heuristic method to automatically choose the hyper-parameter according to the class prior on the training data. Compared with previous approaches, our method only requires a slight modification of the conventional cross-entropy loss and is compatible with various deep neural networks in an end-to-end way. Extensive experiments on synthetic and real-world datasets with and without SCAR assumption verify the superior performance of the proposed method.