Research On Multi-view Clustering Algorithm Based On Graph Learning

Faced with multi-view data with complex attributes from different sources,the performance of traditional clustering methods is limited.The multi-view clustering methods based on graph learning have attracted wide attention because of their simple and effective characteristics.However,these existing multi-view clustering methods based on graph learning usually directly learn the consensus graph of multiple views,which is difficult to accurately express the true structure of all views,thus reducing the clustering accuracy of the algorithm.To solve this problem,on the basis of graph theory,combined with cutting-edge academic ideas such as multi-view learning,local structure preservation and tensor learning,the structural relationship between data samples and different information representation abilities of multiple views are deeply mined,the real cluster structure of data is revealed,and the multiview clustering method based on graph learning is studied and discussed as follows:A robust multi-view clustering method based on latent embedded space learning is proposed to solve the problem of light noise and other adverse factors contained in the original feature.Firstly,a latent space learning module is designed to reduce the impact of the original noise and "dimensional disaster".Then,in the latent embedding space,the affinity graph of the latent embedding representation is learned,and an optimization evaluator is applied to the learned affinity graph to enhance the robustness of the affinity graph.For the optimization problem of the algorithm,an alternate iterative algorithm is used to optimize the objective function.Experimental results show that the performance of the proposed method is obviously better than the latest methods,which proves the effectiveness and superiority of the latent embedded space learning.A multi-view clustering method based on high order similarity learning is proposed to solve the problem that the clustering method based on multi-view embedding space graph learning ignores the indirect relationship between the data samples and fails to fully learn the inherent information of each view.Firstly,a first-order and second-order similarity cooperative learning module is designed,which can excavate the local structural relations and neighboring structural relations among data samples.Then,a higher-order similarity learning module is designed,which can simultaneously learn the affinity graph of each view and the consensus affinity graph of multiple views,and the consensus affinity graph and each affinity graph learn from each other.Finally,an effective alternate iterative algorithm is designed to optimize the objective function,and the computational complexity and convergence of the proposed algorithm are verified theoretically and experimentally.The proposed algorithm is evaluated on several real datasets,and the experimental results verify the effectiveness of high order similarity learning.A multi-view clustering method based on multi-order similarity learning is proposed to solve the problem that clustering method based on multi-view high-dimensional space graph learning is difficult to mine the space consistency of multiple views.Firstly,a first-order and second-order similarity cooperative learning module is designed to excavate the local structural relations and neighboring structural relations among data samples.Then,a tensor based third-order similarity learning module is designed to mine the third-order correlation between multiple views.For the optimization problem of the model,an alternate iterative algorithm was adopted,and the proposed algorithm was analyzed and evaluated from clustering accuracy,parameter sensitivity analysis and convergence analysis.The experimental results verify the effectiveness of multi-order similarity learning for clustering task.A multi-view complete graph clustering method is proposed to solve the problem that the clustering method based on tensor space graph learning does not fully excavate the correlation between multiple views.This method designs a collaborative learning module of first-order similarity,second-order similarity,third-order similarity and interactive learning module between views,and comprehensively learns the multi-order similarity relationship between data samples and the multi-order correlation relationship between views.Among them,the interactive learning between the predefined affinity graph and each affinity graph contribute to learn a complete graph.An alternate iterative algorithm is used to optimize the objective function and its convergence is verified theoretically.In addition,the effectiveness of the proposed algorithm is evaluated in graph comparison experiment,computational complex comparison experiment and parameter sensitivity analysis experiment.Experimental results show that the performance of the proposed algorithm is better than the multi-view clustering method based on tensor space graph learning.In conclusion,starting from the multi-view clustering method based on graph learning,this thesis gradually proposes the multi-view clustering method based on latent embedded space learning,multi-view clustering method based on high order similarity learning,multi-view clustering method based on multi-order similarity learning and multi-view complete graph clustering method.Experiments show that the proposed method can significantly improve the clustering performance and provide theoretical reference for the application of multi-view clustering method based on graph learning.