An improved density peak clustering algorithm

Abstract

Density based clustering is an important clustering approach due to its ability to generate clusters of arbitrary shapes. Among density based clustering algorithms, the density peak (DP) based algorithm is shown to a potential one with some attractive properties. The DP algorithm calculates the local density of each data, and then the distance of each data to its nearest neighbor with higher density. Based on these two measurements, the cluster centers can be isolated from the non-center data. As a result, the cluster centers can be identified relatively easily and the non-center data can be grouped into clusters efficiently. In this paper we study the influence of density kernels on the clustering results and present a new kernel. We also present a new cluster center selection criterion based on distance normalization. Our new algorithm is shown to be effective in experiments on ten datasets.