Research On Multi-view Consistent Clustering Of Graphs

With the gradual advancement of the change in the era of big data and the advantages of describing objects,multi-perspective cluster analysis has gained more research attention.Compared with single-perspective clustering that simply connects all perspective features,multi-perspective clustering that can handle each perspective feature separately can jointly optimize each perspective,so that each perspective feature can be as utilized to improve the final clustering efficiency as possible.The data features of multiple perspectives contain rich and complementary information.Joint optimization of the data features of multiple perspectives can better explore the internal structural relationship of the data.How to better combine the diversity of multi-perspective to explore the complementary information between multi-perspective features,and gradually mine the consistency and inconsistency of multi-perspective data will be the key to optimize and improve the clustering effect of multi-perspective data,which is also the key and difficult point of this thesis.To this end,this thesis proposes a new graph-oriented consistent multi-view sparse clustering framework,which is based on a multi-view sparse representation clustering method to deal with the consistency of multi-view graphs.The method first decomposes the multi-view data into consistency and inconsistency parts;then uses the similarity measurement method and the KNN(K-nearest neighbor)algorithm to decompose and fuse the multi-view data;and then uses the sparse representation to learn the consistency of the multi-view graph Then,the clustering results are obtained by spectral clustering.Compared with the sparse representation,the low-rank representation can better describe the overall structure of the data,and has better robustness to noisy data and outliers.Based on this,this thesis proposes a new graph-oriented consistent multi-view low-rank clustering framework,which combines the consistency,inconsistency and low-rank representations of multi-view graphs to learn the consistency similarity matrix of multi-view graphs,and then uses the spectral Clustering Get clustering results.For the two proposed multi-view clustering frameworks,an alternate iterative optimization algorithm is designed and implemented to solve the objective function respectively.In order to test the availability of the two graph-oriented consistent multi-view clustering methods proposed in this thesis,we compare the two methods proposed in this thesis with six methods in related fields on six real and commonly used multi-view datasets Extensive experiments were carried out.The experimental results demonstrate the availability of the two methods proposed in this thesis.The selected real and commonly used multi-view datasets range from 2 to 7 perspectives,5 to 31 categories,and 169 to 2000 samples,combined with the parameter sensitivity analysis of the two models,the robustness of the method is proved.Through the convergence analysis experiments,it is proved that the two methods proposed in this thesis can quickly reach convergence within a few iterations.To sum up,this thesis proves the effectiveness and advantages of the two algorithms proposed in this thesis through theoretical analysis and experiments.