Multi-view Subspace Clustering Based On Manifold Regularization And Rank Constraint

Clustering is a very important technology in the field of data mining and machine learning.In many practical applications of data mining,data for the same object can be collected from different sources or obtained through different feature extraction methods,such data is called multi-view data.Traditional single-view clustering methods cannot effectively process multiview data,thus many researchers pay attention to multi-view clustering.In order to discover the inner structure of multi-view data and find a common representation of multi-view data,this thesis proposes two multi-view clustering approaches based on subspace learning.The main works of this thesis are as follows:Firstly,this thesis describes the significance of multi-view clustering and the basic framework of several multi-view clustering paradigms.Next,the representative research and their advantages and disadvantages are introduced.Then,this thesis summarizes the use of the Laplacian matrix in manifold learning and analyzes the working principle of traditional singleview subspace clustering.After that,we propose a multi-view clustering method based on manifold regularization and rank constraint(MRMSC).MRMSC extends the single-view subspace learning method to multi-view scenarios by combining subspace learning and manifold regularization.With the benefit of the rank constraint Laplacian matrix,MRMSC can find a common subspace that has a great structure for clustering.Furthermore,in order to compensate for the lack of generalization ability of linear models,a kernelized version of MRMSC called KMRMSC is proposed.For solving MRMSC and KMRMSC,alternating optimization algorithms are developed.Finally,compared with other methods on the synthetic data and the real-world data,the good clustering performances of MRMSC and KMRMSC are fully proved.In summary,the main contributions of this thesis are: 1)This thesis analyzes the characteristics of multi-view data and propose a new multi-view clustering method called MRMSC.MRMSC can preserve the local structure of each view and find a common subspace that is suitable for clustering.2)This thesis makes use of kernel trick and proposed a nonlinear model KMRMSC.KMRMSC has good generalization ability.3)Alternating optimization algorithms are developed to solve MRMSC and KMRMSC.Complexity and parameter sensitivity are discussed.The detailed experiments on synthetic and real-world data demonstrate the effectiveness of the proposed methods compared to benchmark methods.