Investigating Active Positive-Unlabeled Learning with Deep Networks

Abstract

AbstractPositive-unlabeled (PU) learning deals with the binary classification problem when only positive (P) and unlabeled (U) data are available. Recently, many PU learning models have been proposed based on deep networks and become the SOTA of PU learning. Despite the achievements on the model aspect, theoretical analysis and empirical results have shown that the number and quality of positive data can significantly impact learning performance. Active learning is classically used in machine learning to acquire additional high-quality labelled data, however, there are only a few studies on using active learning in deep PU models. This paper investigates the use of active learning in deep PU models. Specifically, this paper studies the uncertainty query strategy for pool-based active learning and show that due to the “large-small-loss” property of deep networks, the query strategy based purely on uncertainty can achieve diversity simultaneously. Empirical results also illustrate the effectiveness of uncertainty-based queries on active PU learning with deep networks.KeywordsPU learningActive learningDeep neural networks