Research On Spectral Clustering Algorithm Based On Nearest Neighbor Graph Analysis

Spectral clustering algorithm is one of the classical clustering algorithms based on spectral graph theory.NJW(Ng-Jordan-Weiss)algorithm in K-way spectral clustering algorithm is widely concerned because of its simple framework.In this algorithm,similarity measure W between samples is established by Euclidean distance,although the measurement method of Euclidean distance is simple and easy to understand,and is used for most data types.However,the Euclidean distance is based on each independent sample,the relationship between samples is obtained directly,and the established sample measurement is also based on the global sample.The correlation between samples is relatively single,which can not adequately describe the local features between samples,making the similarity features described incompletely,and can not fully reflect the relationship between samples.In addition,the K value of traditional K-nearest neighbor algorithm is the empirical value obtained through many experiments,it is difficult to prove that it is the best value,so it is easy to fall into local optimum.In order to solve the above problems,this paper studies the spectral clustering algorithm based on the analysis of nearest neighbor graph.The main work is as follows:Firstly,For dealing with the shortcomings of similarity measure with Euclidean distance,a spectral clustering algorithm based on K-threshold similarity measure is proposed.The algorithm fully refers to the characteristics of each sample area,uses K-threshold to establish the relationship between local samples,and establishes the global connected graph with the shortest path on the basis of these local correlations,then uses local sample standard deviation as the Gaussian kernel.The value.The algorithm overcomes the neglect of the local details of the samples when the Euclidean distance is used to establish the connection.Secondly,aiming at the problem of K value selection in spectral clustering algorithm based on K threshold similarity measure,a clustering algorithm based on custom K value is proposed.In the spectral clustering algorithm based on K-threshold similarity...