Density based cluster extension and dominant sets clustering

Abstract

With the pairwise data similarity matrix as input, dominant sets clustering has been shown to be a promising clustering approach with some nice properties. However, its clustering results are found to be influenced by the similarity parameter used in building the similarity matrix. While histogram equalization transformation of the similarity matrices removes this influence effectively, this transformation causes over-segmentation in the clustering results. In this paper we present a density based cluster extension algorithm to solve the over-segmentation problem. Specifically, we determine the density threshold based on the minimum possible density inside the dominant sets and then add new members into clusters if the density requirement is satisfied. Our algorithm is shown to perform better than the original dominant sets algorithm and also some other state-of-the-art clustering algorithms in data clustering and image segmentation experiments.