Active learning from label proportions via pSVM

Abstract

Learning from label proportions (LLP), in which the training data is divided into different bags and only the proportions of samples belonging to certain categories in each bag are known, has attracted widespread interest in many application scenarios. However, the existing researches ignore the importance of part instances in which the labels can be easily obtained. In this paper, we propose a novel method called active-pSVM, which incorporates active learning into the large margin framework to solve LLP problem. In detail, we first label the most valuable instances that are selected by QUIRE or uncertainty sampling query strategy. Then, the obtained labeled instances and the given unlabeled data are simultaneously used to construct the prediction model under the large margin framework. Extensive experiments on benchmark datasets show that the proposed active-pSVM outperforms the state-of-the-art approaches for solving LLP problem, i.e. InvCal, alter-∝SVM and LLP-NPSVM.