Density Peaks Clustering Based on Candidate Center and Multi Assignment Policies

Density peak clustering (DPC) the center selection ignores the environment of samples, which can easily result in the incorrect identification of cluster centers. Additionally, a single allocation approach can easily produce the joint effect of allocation mistakes. To address the above issues, a density peaks clustering based on candidate center and multi assignment policies (DPC-CM) is proposed. Firstly, to retain cluster centers with relatively small densities, DPC-CM introduces thresholds to delineate candidate center regions in the decision graph. It also incorporates shared nearest neighbor definitions to ensure proper cluster center selection. Then, the allocation results of the cluster backbone are corrected using shared nearest neighbor information, suppressing allocation cascading errors effectively. Finally, the allocation of cluster backbone points used as the basis for label propagation also improves the allocation efficiency while taking into account the surroundings in which the sample sites are situated. The fault tolerance of the algorithm has also been enhanced by the multi-step allocation strategy. Experimental results on synthetic and UCI datasets validate that the DPC-CM algorithm outperforms comparative algorithms for clustering on complex morphological and unevenly distributed datasets.