Research and development of spectral clustering algorithms

Abstract

Spectral clustering is a very competitive clustering method. It attracts more and more attention of academia in recent years and becomes a new hotspot in machine learning since it does not make any assumptions on the global structure of data and has excellent performance for the sample space of arbitrary shape. The idea of spectral clustering is based on spectral graph theory. It treats data clustering problem as a graph partitioning problem and we can get the best clustering results by constructing a suitable graph, and using the appropriate graph cut method. Spectral clustering is better than traditional clustering algorithms in many aspects and has been successfully applied in areas such as data analysis, speech and image processing. This paper first introduces the basic concepts of graph theory and traditional graph cut methods; then analyses typical spectral clustering algorithms and reviews the latest development of spectral clustering; finally, proposes several valuable research directions in light of the deficiencies of spectral clustering algorithms.