An improved density peaks clustering algorithm using similarity assignment strategy with K-nearest neighbors

Some particular shaped datasets, such as manifold datasets, have restrictions on density peak clustering (DPC) performance. The main reason of variations in sample densities between clusters of data and uneven densities is not taken into consideration by the DPC algorithm, which could result in the wrong clustering center selection. Additionally, the use of single assignment method is leads to the domino effect of assignment errors. To address these problems, this paper creates a new, improved density peaks clustering method use the similarity assignment strategy with K nearest Neighbors (IDPC-SKNN). Firstly, a new method for defining local density is proposed. Local density is comprehensively consider in the proportion of the average density inside the region, which realize the precise location of low-density clusters. Then, using the samples' K-nearest neighbors information, a new similarity allocation method is proposed. Allocation strategy successfully address assignment cascading mistakes and improves algorithms robustness. Finally, based on four evaluation indicators, our algorithm outperforms all the comparative clustering algorithm, according to experiments conducted on synthetic, real world and the Olivetti Faces datasets.