Density peaks clustering based on k-nearest neighbors sharing

The density peaks clustering (DPC) algorithm is a density-based clustering algorithm. Its density peak depends on the density-distance model to determine it. The definition of local density for samples used in DPC algorithm only considers distance between samples, while the environments of samples are neglected. This leads to the result that DPC algorithm performs poorly on complex data sets with large difference in density, flow pattern or cross-winding. In the meantime, the fault tolerance of allocation strategy for samples is relatively poor. Based on the findings, this article proposes a density peaks clustering based on k-nearest neighbors sharing (DPC-KNNS) algorithm, which uses the similarity between shared neighbors and natural neighbors to define the local density of samples and the allocation. Comparison between theoretical analysis and experiments on various synthetic and real data reveal that the algorithm proposed in this article can discover the cluster center of complex data sets with large difference in density, flow pattern or cross-winding. It can also provide effective clustering.

Automated summary

The DPC algorithm performs poorly on complex data sets with large differences in density, flow pattern or cross-winding, and has relatively poor fault tolerance in sample allocation. This article proposes the DPC-KNNS algorithm to improve clustering performance on datasets with large density differences and complex patterns.