An Automatic Index Validity for Clustering

Many validity index algorithms have been proposed to determine the number of clusters. These methods usually employ the Euclidean distance as the measurement. However, it is difficult for the Euclidean distance metric to evaluate the compactness of data when non-linear relationship exists between different components of data. Moreover, most current algorithms can not estimate well the scope of the number of clusters. To address these problems, in this paper, we adopt the kernel-induced distance to measure the relationship among data points. We first estimate the upper bound of the number of clusters to effectively reduce iteration time of validity index algorithm. Then, to determine the number of clusters, we design a kernelized validity index algorithm to automatically determine the optimal number of clusters. Experiments show that the proposed approach can obtain promising results.