Research On Robust Spectral Clustering Method Based On Constrained Laplacian Rank

With the development of the information technology and the internet of things,big data has penetrated into every field of industry and business.However,due to its large quantity and low value density,people make use of data inefficiently.Whereupon how to analyze and mine available information from data has important academic and application value,and thus become a hot and difficult research topic.Cluster analysis is the key to probe data analysis,which can be widely used in image segmentation and unsupervised classification due to its ability to group data according to the information between the data without any prior information.In recent years,utilizing the clustering methods of graph-based for image segmentation or classification has received widespread attention for their satisfactory performance,and how to better use graph represent the relationship between data has become the pivotal part of the task.The paper proposes a new clustering algorithm model via robust graph learning,which is based on the idea of spectral clustering.First,the noise in the original data is separated by the method of robust principal component analysis,then the similar graph that can represent the similarity connection of data can be constructed according to the knowledge of graph theory by using the clean data processed,after then,this model imposes the Laplacian rank constraint on the graph to complete the classification task.This paper also designs an optimization solution algorithm to deal with the new proposed model and gives the specific process.In addition,this paper comprehensively considers the associated information between views that constitute the multi-view and sharing weighting of cluster,and proposes a new multi-view clustering model based on the robust graph learning model.This model adopts a parameter weighting method to assign the reasonable weight to each similar graph and integrates different similar graphs by learning.In order to verify the feasibility and superiority of the proposed two models,experiments are conducted on many image datasets including public image datasets and a self-built design patent dataset with some result comparison to related current mainstream methods.Experimental results indicate that both the single-view and multi-view algorithm models proposed in this paper have better clustering performance,which effectively improve the accuracy of clustering.Through comparison,it can be shown that the proposed methods have better overall performance and better robustness,and can handle the tasks of image segmentation or unsupervised data classification more effectively.