Sampling Active Learning Based on Non-parallel Support Vector Machines

Abstract

Labeled examples are scarce while there are numerous unlabeled examples in real-world. Manual labeling these unlabeled examples is often expensive and inefficient. Active learning paradigm seeks to handle this problem by identifying the most informative examples from the unlabeled examples to label. In this paper, we present two novel active learning approaches based on non-parallel support vector machines and twin support vector machines which adopt the margin sampling method and the manifold-preserving graph reduction algorithm to select the most informative examples. The manifold-preserving graph reduction is a sparse subset selecting algorithm which exploits the structural space connectivity and spatial diversity among examples. In each iteration, an active learner draws the informative and representative candidates from the subset instead of the whole unlabeled data. This strategy can keep the manifold structure and reduce noisy points and outliers in the whole unlabeled data. Experimental results on multiple datasets validate the effective performance of the proposed methods.