Study on Arbitrary Distribution in Cluster Analysis

Three clustering methods are presented and discussed by experimental analysis. The results by using three clustering methods which are partitioning methods, hierarchical methods and density-based methods visually illustrate the clustering results, in two-dimensional data sets as experimental data are used. Clearly, when the original data set is spherical shape, most of the cluster methods can get good clustering results. Partitioning methods (K-means) can´t handle clusters of arbitrary shapes and different sizes, and can´t handle clusters of varying densities. Hierarchical methods can identify globular clusters well whether globular clusters is in same densities or in varying densities, and this approach can handle clusters of winged shapes of well departed, but cannot handle clusters of winged-globular shapes. Based on the notions of density and density reachable, the CADD (clustering algorithm based on object density and density-reachable) can find clusters of arbitrary shapes and different sizes, and can handle clusters of varying densities.