Clustering Based on Dominant Set and Cluster Expansion

Abstract

While numerous clustering algorithms can be found in the literature, existing algorithms are usually afflicted by two major problems. First, the majority of clustering algorithms requires user-specified parameters as input, and their clustering results rely heavily on these parameters. Second, many algorithms generate clusters of only spherical shapes. In this paper we try to solve these two problems based on dominant set and cluster expansion. We firstly use a modified dominant sets clustering algorithm to generate initial clusters which are parameter independent and usually smaller than the real clusters. Then we expand the initial clusters based on two density based clustering algorithms to generate clusters of arbitrary shapes. In experiments on various datasets our algorithm outperforms the original dominant sets algorithm and several other algorithms. It is also shown to be effective in image segmentation experiments.