A multiseed non-hierarchical clustering technique for data analysis

Abstract

Clustering methods such as K-means and its variations, such as Forgy, as well as their improved version ISODATA, do not work well if the shape of the cluster is elongated. It is pointed out that a single seed point cannot correctly reflect the nature of the data of an elongated cluster. A multiseed clustering algorithm is proposed, where one cluster may contain more than one seed point. A density-based algorithm is used to choose the initial seed points. To assign several seed points to one cluster, a minimal spanning tree guided novel merging technique is proposed. The merging technique is quite general and may be applied to other clustering approaches as well. Experimental results are presented to demonstrate the efficiency of this clustering procedure