Particle Swarm Optimization Based Clustering: A Comparison of Different Cluster Validity Indices

Most of clustering algorithms based on natural computation aim to find the proper partition of data to be processed by optimizing certain criteria, so called as cluster validity index, which must be effective and can reflect a similarity measure among objects properly. Up to now, four typical cluster validity indices such as Euclid distance-based PBM index, the kernel function induced CS measure, Point Symmetry (PS) distance-based index, Manifold Distance (MD) induced index have been proposed. But, there is not a detailed comparison among these indexes. In this paper, we design a particle swarm optimization based clustering algorithm, in which, four different cluster validity index above mentioned are used as the fitness of a particle respectively. By applying the proposed algorithm to a number of artificial synthesized data and UCI data, the performance of different validity indices are compared in terms of clustering accuracy and robustness at length.