A projection-based split-and-merge clustering algorithm. (February 2019)

• Record Type: Journal Article
• Authors: Cheng, Mingchang; Ma, Tiefeng; Liu, Youbo
• Is Part Of: Expert systems with applications. Volume 116(2019)
• Page Start: 121
• Page End: 130
• Publication Date: 2019-02
• Subjects: Split-and-merge -- Projection -- K-means -- Density curve; 68T10; Expert systems (Computer science) -- Periodicals; Systèmes experts (Informatique) -- Périodiques; Electronic journals; 006.33
• DOI: 10.1016/j.eswa.2018.09.018
• Languages: English
• ISSNs: 0957-4174

Abstract

Highlights: Sufficient dimension-reduction through projecting sample points on a line. Split or merge clusters according to the density curve of the projection points. Use split-and-merge process to adaptively give a certain number of clusters. Abstract: A novel split-and-merge clustering algorithm is proposed by using projection technology and K-means method. There are two key technologies in the proposed method: shape recognition based on projection and split-and-merge process based on K-means. By projecting the data onto the connection of any two cluster centers, no matter how large the dimension of data is, we can always obtain an one-dimension density curve of the projection to guarantee an acceptable amount of calculation. Further embedded the kernel density estimate, we can determine the distribution of clusters by the shape of the one-dimensional density curve. In the split-and-merge process, this algorithm not only addresses the sensitivity in selecting initial cluster centers, but also automatically give a reasonable number of clusters. We also discuss the possibility to extend the projection split-and-merge method from K-means to density based methods (as EM algorithm and Cross-entropy clustering). Both simulation and real data experimental results show that our method performance well especially under strict data conditions.