Density Peak Clustering Algorithm Based on Optimal Density Radius

Abstract

The density peaks clustering (DPC) algorithm is one of the most important progress in recent clustering algorithms, which needs neither any iterative process nor more parameters, and thus takes advantages over most existing clustering algorithms. But the density radius is an uncertain parameter in DPC, and its different values may lead to very different clustering results. This problem greatly limits its applicable range. In this paper, an efficient method is proposed to determine the density radius. The core idea is that an optimal density radius must maximize the density differences of all samples. Consequently, the uncertain parameter in the DPC algorithm is optimally determined. The experimental results of a set of real data sets with different structures show that the improved DPC algorithm has higher clustering accuracy than the original DPC algorithm, and essentially has more robust clustering results.