A robust density peaks clustering algorithm with density-sensitive similarity

Abstract

Density peaks clustering (DPC) algorithm is proposed to identify the cluster centers quickly by drawing a decision-graph without any prior knowledge. Meanwhile, DPC obtains arbitrary clusters with fewer parameters and no iteration. However, DPC has some shortcomings to be addressed before it is widely applied. Firstly, DPC is not suitable for manifold datasets because these datasets have multiple density peaks in one cluster. Secondly, the cut-off distance parameter has a great influence on the algorithm, especially on small-scale datasets. Thirdly, the method of decision-graph will cause uncertain cluster centers, which leads to wrong clustering. To address these issues, we propose a robust density peaks clustering algorithm with density-sensitive similarity (RDPC-DSS) to find accurate cluster centers on the manifold datasets. With density-sensitive similarity, the influence of the parameters on the clustering results is reduced. In addition, a novel density clustering index (DCI) instead of the decision-graph is designed to automatically determine the number of cluster centers. Extensive experimental results show that RDPC-DSS outperforms DPC and other state-of-the-art algorithms on the manifold datasets.