关键词: 聚类算法, 谱聚类算法, [K]-均值算法, 相似矩阵

Abstract:

Cluster analysis is a common analysis method. As a branch of cluster analysis, spectral clustering has attracted much attention because of its characteristics such as not being restricted by sample shape. In order to timely grasp the current research trends of spectral clustering algorithm, the spectral clustering optimization algorithms are divided into three categories from three perspectives：semi-supervised learning, two-stage clustering algorithm selection and algorithm execution efficiency optimization, and the optimization ideas of each category of algorithms are summarized. Firstly, the classical k-way spectral clustering and its basic theory are introduced, the reasons and influences of the selection of similarity matrix, eigenvalues and eigenvectors are introduced and analyzed. The purpose is to clarify its importance in the clustering process and the necessity of optimization in this part. Furthermore, based on the difference of algorithm improvement strategies, it sorts out and summarizes the improvement ideas, research status, advantages and disadvantages of each type of algorithm. Finally, the spectral clustering algorithm and optimization algorithms are compared on UCI data sets and handwritten data sets, and the future research trends in spectral clustering optimization algorithm are discussed.

Key words: clustering algorithm, spectral clustering;[K]-means, affinity matrix