A Support Vector Clustering Type Algorithm via Minimum Enclosing Balls

• Main Authors: Hsi-Chen Tsai, 蔡錫震
• Other Authors: Yuh-Jye Lee
• Format: Others
• Language: en_US
• Published: 2006
• Online Access: http://ndltd.ncl.edu.tw/handle/da7yaa

碩士 === 國立臺灣科技大學 === 資訊工程系 === 94 === Clustering analysis which is categorized as unsupervised learning in machine learning means based on speci‾c features creating groups of objects in such a way that the objects grouping into the same clusters are similar and those belonging in diferent clusters are dissimilar. Support vector clustering (SVC) is an unsupervised and kernel-based clustering algorithm. SVC could naturally separate dataset with any shape into diferent clusters. SVC separates the dataset into appropriate clusters by tuning two parameters. Suppose number of data points is n, the time complexity of labeling data points becomes O(n2). It becomes the bottleneck of SVC and this is the major drawback why SVC always takes more time than other clustering algorithms. Focus this drawback, this thesis suggests a novel cluster labeling algorithm combining with the concept of minimum enclosing balls (MEBs), the property of support vector and k means to improve the e±ciency. In the later section, we will test our proposed method on synthetic datasets either in R2 and R3 space in order to visualize the clustering results, and demonstrate the e±ciency of our proposed method by comparing with other cluster labeling algorithms. Besides we also find a flaw in the original SVC mathematical programming model, another issue of this thesis is to discuss the flaw and solve it.