Selection of gaussian kernel widths and fast cluster labeling for support vector clustering
Sei-Hyung Lee, Karen Daniels · 2005
Clustering forms natural groupings of data points that maximize intra-cluster similarity and minimize inter-cluster similarity. Support Vector Clustering (SVC) is a clustering algorithm which can handle arbitrary shapes of clusters and outliers. The two most important input parameters of SVC, when using a Gaussian kernel, are the width of the Gaussian kernel and a soft margin constraint to control outliers. These parameters control the number of clusters. Two main problems of SVC are choosing appropriate values for the parameters and assigning cluster labels to data points. Since we assume that we have no prior knowledge about the data set, finding appropriate parameter values becomes a difficult problem. This dissertation suggests a method to explore some appropriate ranges of the Gaussian kernel width values. Almost all clustering algorithms require some parameters and their performance depends on the selection of parameter values. Our parameter value selection method for SVC generates several good values to obtain reasonable clusterings. Since the cluster labeling process is expensive, we propose a new novel cluster labeling algorithm to improve efficiency without losing accuracy of clustering. We present comparison results of our cluster labeling algorithm and other cluster labeling algorithms. The results show that our approach is fast compared to those traditional cluster labeling algorithms.