Subordinate based Cluster Center Identification in Density Peak Clustering

Abstract

Recently, a clustering algorithm is proposed by treating local density peaks as cluster centers. This algorithm proposes to describe the data to be clustered with local density and the distance of one data to the nearest data of larger local density. This description highlights the uniqueness of cluster centers and is utilized to determine cluster centers. With the assumption that one data and the nearest data of larger local density are in the same cluster, the non-center data are assigned labels efficiently. By studying the clustering process of this algorithm in depth, we find that the local density is not very effective in highlighting the uniqueness of cluster centers. As a result, this algorithm is dependent on the parameters in local density calculation. We discuss this problem and find that it is the role of density peaks, but not the absolute local density, that highlights the uniqueness of cluster centers. Based on this observation, we introduce the concept of subordinate and use the amount of subordinates to replace the local density in cluster center identification. Together with a new density kernel, this new criterion is shown to be effective in experiments and comparisons.