CSBD: A Nonlinear Clustering Method Based on Cluster Shrinking and Border Detection

Abstract

Nonlinear clustering has attracted an increasing amount of attention recently. In this paper, we propose a new nonlinear clustering method based on Cluster Shrinking and Border Detection (CSBD). Unlike most existing clustering method, the CSBD method focuses on every data point rather then the cluster centers. A novel idea, namely Cluster Shrinking, is designed to transform the original nonlinear datasets into several hyperspheres, which makes clustering work much easier. Besides, we also introduce a simple but effective Border Detection method based on histogram analysis to automatically determine the threshold parameter in Cluster Shrinking phase. Extensive experiments have been conducted to demonstrate the effectiveness of CSBD in both synthetic and real-world datasets.