Research And Application Of Spectral Clustering Based On Density Adaptive Neighborhood

Spectral clustering has got more focus in research recently.Because it shows a predominant capability in processing the non-convex spatial data.Similarity matrix is the basis to ensure the success of the spectral clustering.The existing spectral clustering algorithms are generally based on the undirected K nearest neighbor graph and Gaussian kernel function to construct the similarity matrix,which cannot well reveal the part of the data.And when constructing the similarity matrix,it relies on empirical values to determine the appropriate parameters,which is very sensitive to the selection of parameters.This thesis studies the issues above,and the specific research content is as follows:Firstly,this thesis proposes a density-adaptive neighborhood spectral clustering based on shared nearest neighbors.Aiming at the problem that spectral clustering for datasets with complex shapes and different density,it is difficult to determine the appropriate construction of similarity graphs without prior information,and the similarity measure of Gaussian kernel function based on Euclidean distance ignores global consistency.In this thesis,a density-adaptive neighborhood construction algorithm is used to construct neighborhood information without parameters,and the shared nearest neighbor is used as a measure of the similarity between samples.In this way,the influence of parameters in the construction of similar graphs is eliminated,and local density information is reflected.The experiments certify that the proposed algorithm has higher precision compared to the K-means and the spectral clustering based on K nearest neighbors.Meanwhile it is insensitive to parameter selection,which facilitates the spectral clustering algorithm applied in different fields without priori information.Secondly,this thesis proposes a density adaptive neighborhood spectral clustering algorithm based on selfish herd optimization.When the similarity matrix is constructed,the spectral clustering obtains the vector space mapping the original data through Laplace's eigenvector decomposition,and finally uses the K-means algorithm to cluster it.However,the K-means algorithm is sensitive to the random initialization of clustering centers,which brings instability to the overall spectral clustering algorithm.When faced with data with a large number of categories,it is easy to produce unsatisfactory clustering results.The selfish herd optimization algorithm is introduced into the K-means,and the optimal initial clustering center is solved through the optimization algorithm.Thereby,the performance of spectral clustering is improved.The experiments certify that the clustering effect of the proposed algorithm is stable,and it has good clustering results for data sets with different numbers of categories.Finally,this thesis explains the key technology and implementation process of text clustering.Then the proposed algorithm is applied to the field of text clustering.Through text clustering experiments on real text data,the results reveal that the proposed algorithm can well solve the problem of text clustering in different categories.It is feasible in solving practical problems.