¹College of Computer Science and Technology, Xuzhou Normal University, Xuzhou, China
²Training Department, Xuzhou Air Force College of PLA, Xuzhou, China
³Department of Logistic Command, Xuzhou Air Force College of PLA, Xuzhou, China

---

Abstract—As an important issue of machine learning, clustering receives much care in recent years. Among all clustering approaches, most of them conduct clustering operations on overall data. That is, they learn label information from all data. That comes across critical challenge in times of high-sized datasets. This paper proposes a novel Three-phase Labeling algorithm (TPL) based on SVC to overcome this problem. TPL consists of selecting data representatives (Data representatives), clustering (Data representatives) and then classifying non- Data representatives respectively. Support vector clustering process is modified to select qualified Data representatives in first phase. Spectrum technique governs the second-phase clustering task. Therein, the geometric properties of feature space, a new metric, and a tuning strategy of Kernel scale are used. In experiments on real datasets, TPL achieves clear improvement in accuracy and efficiency over its counterparts, and demonstrates highly competitive clustering performance in comparison with some state of the arts.

Index Terms—Three-phase clustering, support vector clustering, data representatives, new metric, tuning strategy.

[PDF]

Cite: Ping Ling, Xiangsheng Rong, Xiangyang You, Ming Xu, "Novel Three-Phase Clustering based on Support Vector Technique," Journal of Software vol. 8, no. 4, pp. 955-962, 2013.