Density Peaks Clustering Based on Potential Model and Diffusion Strength

Abstract

Density peaks clustering (DPC) is a simple and efficient density-based clustering algorithm without complex iterative procedures. However, DPC needs to manually choose clustering centers via a decision graph, which often can't identify real centers and breaks the continuous flow of the algorithm. In addition, DPC is highly sensitive to the cut-off distance and suffers from the domino chain reaction. To surmount the aforementioned deficiencies, an improved density peaks clustering based on potential model and diffusion strength (DPC-PMDS) is proposed in this paper. Firstly, we utilize the potential and centrality of data points to calculate the density instead of the cut-off distance. Secondly, inspired by the information diffusion in social networks, we define the influence of data points and the diffusion strength between data points, and realize the diffusion of label from each center to the core data points while selecting clustering centers. Through this process, the core structure of each clustering is obtained and the centers are accurately identified. Finally, the distances from the boundary points to each cluster computed based on centrality are applied to assign boundary points to avoid chain reaction. Extensive experiments on synthetic, UCI and Olivetti Faces datasets demonstrate that DPC-PMDS can achieve excellent clustering results over other state-of-art algorithms, especially on datasets with complex shapes and uneven density distribution.