Experimental evaluation of a density kernel in clustering

Abstract

The recently proposed clustering algorithm based on density peaks is reported to generate very good clustering results. This algorithm is simple and efficient, and can be used to generate clusters of arbitrary shapes. However, the performance of this algorithm relies on the selection of the kernel in local density calculation. The original density peak based algorithm uses the cutoff kernel and Gaussian kernel to calculate the local density, and the clustering results are found to be influenced by the cutoff distance, which can only be determined empirically so far. In this paper we use a different kernel in density calculation, and evaluate the influence of related parameter on the clustering results. Our work is helpful in understanding the clustering mechanism of this algorithm.