An active learning-based SVM multi-class classification model. Issue 5 (May 2015)

• Record Type: Journal Article
• Authors: Guo, Husheng; Wang, Wenjian
• Is Part Of: Pattern recognition. Volume 48:Issue 5(2015:May)
• Page Start: 1577
• Page End: 1597
• Publication Date: 2015-05
• Subjects: Multi-class classification with unknown categories -- Active learning -- Support vector machine -- MC_SVMA model; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2014.12.009
• Languages: English
• ISSNs: 0031-3203

Abstract

Abstract: Traditional multi-class classification models are based on labeled data and are not applicable to unlabeled data. To overcome this limitation, this paper presents a multi-class classification model that is based on active learning and support vector machines (MC_SVMA), which can be used to address unlabeled data. Firstly, a number of unlabeled samples are selected as the most valuable samples using the active learning technique. And then, the model quickly mines the pattern classes for unlabeled samples by computing the differences between the unlabeled and labeled samples. Moreover, to label the unlabeled samples accurately and acquire more class information, the active learning strategy is also used to select compatible, rejected and uncertain samples, which are labeled by experts. Thus, the proposed model can determine as many classes as possible while requiring fewer samples to be manually labeled. This approach permits an unlabeled multi-classification problem to be translated into a classical supervised multi-classification problem. The experimental results demonstrate that the MC_SVMA model is efficient and exhibits good generalization performance. Highlights: The proposed MC_SVMA model is a deep fusion of active learning and multi-classification. The MC_SVMA model can mine categories from given unlabeled samples quickly with much less labeling cost. Those difficult distinguished unlabeled samples can be classified and new categories may be found. The MC_SVMA model can obtain fast learning and good performance for unlabeled multiple classification problems.