Density Peaks Clustering Algorithm Based on K Nearest Neighbors

Abstract

Density peaks clustering algorithms calculate the local density based on the cutoff distance and the global distribution of the sample. They cannot capture the local characteristics of the sample well, and are prone to appear errors in the selection of density peaks; additionally, the allocation strategy has poor fault tolerance. Once a sample is allocated incorrectly, subsequent allocations will magnify the error. Hence, we proposed a density peaks clustering algorithm based on k-nearest neighbors (DPC-KNN). First, the k-nearest neighbors information of the sample is used to define the local density of the sample in order to find the cluster centers accordingly; the sample with the distance between cluster centers and k-nearest neighbors sample less than the set threshold is defined as the core sample, and the core sample is classified into the corresponding cluster to construct the core area of the cluster; after the degree of attribution of the remaining samples and various clusters are calculated, they are allocated to clusters with high degree of attribution. In order to verify the effectiveness of the proposed algorithm, eight synthetic datasets and ten UCI datasets are selected for experiments, and the proposed algorithm is compared with FKNN-DPC, DPCSA, FNDPC, DPC and DBSCAN. The experimental results indicated that the proposed algorithm had better clustering performance.