A comparative study of different cluster validity indexes

The cluster validity index plays an important role in most clustering algorithm based natural computations. So far, four typical cluster validity indexes have been proposed for clustering data with different structures, including the Euclid distance based Pakhira-Bandyopadhyay-Maulik index, the kernel function induced Chou-Su measure, the point symmetry distance based index and the manifold distance (MD) induced index. However, there is no detailed comparison made among these indexes. This paper compares these four cluster validity indexes by using a simple clustering technique based on particle swarm optimization (PSO). Extensive experiments on a large number of artificial synthesized data sets and UC Irvine data sets, texture images and synthetic-aperture radar images are performed in order to make a comprehensive comparison. Experimental results show that the PSO-based clustering algorithm using the MD induced index has a good performance on most of the data sets.