A Robust Density Clustering Algorithm Based on Gravity Peak

Abstract

Density Peak clustering algorithm has become a hotspot of clustering research in recent years because of its own characteristics of finding high-density centers. However, it and its derivatives are generally not good at distinguishing the cluster centers of uniform and heterogeneous data sets. Moreover, the common problem of the density-based clustering algorithm is that it cannot effectively distinguish clusters with overlaps. To address those problems, a new robust density clustering algorithm called RDC-GP is proposed in this paper. Introducing the law of gravity into density-based clustering algorithm can solve the first problem. In addition, we propose a gravity tilt detection method to identify and extract the edge points of clusters. It not only improve the accuracy of density clustering algorithm, but also well solve the defect that density peak algorithm cannot find cluster centers in unbalanced data effectively. We selected various types of data sets for experiments. Compared with other methods, our proposed has obvious robustness and significantly accuracy.