FRSVC: Towards making support vector clustering consume less. (September 2017)

• Record Type: Journal Article
• Authors: Ping, Yuan; Tian, Yingjie; Guo, Chun; Wang, Baocang; Yang, Yuehua
• Is Part Of: Pattern recognition. Volume 69(2017:Sep.)
• Page Start: 286
• Page End: 298
• Publication Date: 2017-09
• Subjects: Large-scale data -- Support vector clustering -- Dual coordinate descent -- Sampling strategy; Pattern perception -- Periodicals; Perception des structures -- Périodiques; Patroonherkenning; 006.4
• DOI: 10.1016/j.patcog.2017.04.025
• Languages: English
• ISSNs: 0031-3203

Abstract

Highlights: Pricey storage and computation consumptions frustrate SVC's application on limited platforms. We propose FRSVC: towards making SVC consume less and affordable for any platform. A reformative solver for dual problem is proposed based on dual coordinate descent method. Sample once with connection checking first strategy is designed for a faster labeling phase. Results confirm the superiority of FRSVC under limited-memory constraints. Abstract: In spite of with great advantage of discovering arbitrary shapes of clusters, support vector clustering (SVC) is frustrated by large-scale data, especially on resource limited platform. It is due to pricey storage and computation consumptions from solving dual problem and labeling clusters upon the pre-computed kernel matrix and sampling point pairs, respectively. Towards on it, we first present a dual coordinate descent method to reformulate the solver that leads to a flexible training phase carried out on any runtime platform with/without sufficient memory. Then, a novel labeling phase who does connectivity analysis between two nearest neighboring decomposed convex hulls referring to clusters is proposed, in which a new designed strategy namely sample once connected checking first tries to reduces the scope of sampling analysis. By integrating them together, a faster and reformulated SVC (FRSVC) is created with less consumption achieved according to comparative analysis of time and space complexities. Furthermore, experimental results confirm a significant improvement on flexibility of selective efficiency without losing accuracy, with which a balance can be easily reached on the basis of resources a platform equipped.