A non-parametric density kernel in density peak based clustering

Abstract

Density peak (DP) based clustering algorithm is a recently proposed clustering approach and has been shown to be with great potential. This algorithm is based on the simple assumption that cluster centers have high local density and they are relatively far from each other. This observation is used to isolate cluster centers from other data. By making use of the density relationship among neighboring data, the non-center data can be grouped into their respective clusters efficiently. Local density calculation is a key step of the DP algorithm and it influences the intermediate calculation and final result. Existing density kernels usually involve parameters which impact on density values and clustering results. In this paper we present a non-parametric density kernel based on the properties of the dominant sets algorithm. We firstly show that the weights of data in a dominant set can be used as a measure of local density. Then an efficient method is proposed to reduce the computation load. We use experiments on different types of datasets and comparison with other algorithms demonstrate the effectiveness of the proposed approach.