Information-Theoretic Representation Learning for Positive-Unlabeled Classification

摘要

Recent advances in weakly supervised classification allow us to traina classifier from only positive and unlabeled (PU) data. However, existingPU classification methods typically require an accurate estimateof the class-prior probability, a critical bottleneck particularly for highdimensionaldata. This problem has been commonly addressed by applyingprincipal component analysis in advance, but such unsuperviseddimension reduction can collapse the underlying class structure. In thisletter, we propose a novel representation learning method from PU databased on the information-maximization principle. Our method does notrequire class-prior estimation and thus can be used as a preprocessingmethod for PU classification. Through experiments, we demonstrate thatour method, combined with deep neural networks, highly improves theaccuracy of PU class-prior estimation, leading to state-of-the-art PU classificationperformance.