Density peaks clustering based on density voting and neighborhood diffusion

Density Peaks Clustering (DPC) is a well-known clustering technique in the data mining field with fewer parameters as well as no iteration. However, when dealing with datasets containing multiple peaks, DPC may subjectively choose the wrong cluster centers through the decision graph. Additionally, DPC requires a considerable amount of time to estimate density and relative distance. Moreover, DPC is sensitive to the value of cut-off distance. To overcome these issues, a density peaks clustering algorithm based on density voting and neighborhood diffusion (DPC-DVND) is proposed. Firstly, the proposed algorithm utilizes the k nearest neighbors and KD-tree to enhance the efficiency of computing local density and relative distance. Secondly, this study selects the potential cluster centers by density voting and applies the number of votes instead of density to calculate the feasibility of each potential center becoming a cluster center, so that the centers of low-density clusters can be better distinguished. Finally, two neighborhood density diffusion rules are designed to propagate labels and form the core structure of clusters. Experiments on synthetic, real, and image datasets are performed to compare different methods. Results show that DPC-DVND outperforms other state-of-the-art algorithms in terms of effectiveness and efficiency.