Local density peaks clustering with small size distance matrix

Abstract

Clustering by using density peaks (DPC) is a powerful clustering method. The cluster centers in DPC are characterized by a higher density than their neighbors and by a relatively large distance from points with higher densities, which leads to the storage of a distance matrix with the same size of sample size. In this paper, to deduce the storage of the distance matrix, we propose a local density peaks clustering (LDPC) by introducing the local region by reverse nearest neighbors. Compared with DPC, the potential cluster centers and the potential haloes are defined, and the potential cluster centers distance matrix is introduced to replace the storage of the sample distance matrix, which makes it more practical for large scale data. At the same time, the local information is used to construct the high densities, leads to more robust to the data structure. Further, the second order differential operator is used to give the number of clustering automatically. Preliminary computational results demonstrate the effectiveness of proposed LDPC over DPC on artificial datasets and face image considered.