Star-Scan: A Stable Clustering by Statistically Finding Centers and Noises

Abstract

In this paper, we present a new clustering algorithm, called A $\mathbf{Sta }$ble Cluste$\mathbf{r }$ing by $\mathbf{S }$tatistically Finding $\mathbf{C }$enters $\mathbf{a }$nd $\mathbf{N }$oises (Star-Scan). Star-Scan is a density-based clustering algorithm that can find arbitrary shape clusters and resists to the noise in a dataset. It borrows the idea from Rodriguez’s Clustering by Fast Search and Find of Density Peaks (CFSFDP) that the cluster centers are characterized by the points with both higher density and farther distance to other centers than their neighbors. Different from CFSFDP, instead of manual operation, Star-Scan uses a statistical method, box plot, to select cluster centers automatically. Furthermore, due to inadequate selection of cluster centers in CFSFDP, we apply a merging post-process to the produced clusters to get stable and correct results. Finally, we also use box plot to filter out noises on each of final clusters to solve the problem of over-filtering in CFSFDP. We have demonstrated the good performance of Star-Scan algorithm on several synthetic datasets.