PDCSN: A partition density clustering with self-adaptive neighborhoods

Density-based clustering can discover convex and non-convex clusters without specifying the number of clusters. However, its ability to handle clusters with heterogeneous densities is limited. Although various variants solve this problem to some extent, they are still powerless for adjacent clusters with similar densities. Furthermore, the clustering performance is heavily dependent on user-specified parameters. This paper proposes a partition density clustering with self-adaptive neighborhoods (PDCSN). For the parameter dependence problem, a self-adaptive approach based on the natural neighborhood is designed. This approach utilizes the differences and intrinsic properties of sample density distributions to automatically find the optimal neighborhood size A. A partitioning strategy based on mutual A-nearest neighbors-connectedness is proposed to distinguish clusters with large varying densities. Moreover, a core sample search approach based on shared similarity and density heterogeneity is proposed to identify adjacent clusters with similar densities. A series of experiments on 9 synthetic, 10 real-world, and 2 image datasets demonstrate that PDCSN outperforms several famous clustering algorithms.