I-CFSFDP: A Robust and High Accuracy Clustering Method Based on CFSFDP

Clustering by fast search and find of density peaks (CFSFDP) is honored by its simplicity and speed. However, the hyper-parameter $d_{c}$ can only be determined through empirical experience, and for those datasets containing points that are closer to other cluster center, the algorithm does not perform well. In this work, we proposed a new method called Improved CFSFDP (I-CFSFDP), which makes two modifications compared to CFSFDP. Firstly, a new indicator for density is introduced to eliminate the effect of $d_{c}$ on clustering results. Secondly, cluster diffusion model was proposed to cluster remaining points after finding cluster centers. When regarding datasets as graphs, this process can be abstracted as finding the minimum spanning forest model in a graph, and each spanning tree represents a cluster in the dataset. I-CFSFDP is comprehensively evaluated on several datasets with arbitrary distribution and has demonstrated that I-CFSFDP is distinctly more accurate and robust than CFSFDP and DBSCAN.