Research On Multi-View Clustering Algorithm Based On Subspace Learning

Multi-view clustering aims to partition multi-view data into several clusters according to their underlying distribution structures.Subspace learning based multi-view clustering models the structures as multiple linear subspaces and addresses the clustering problem in subspace segmentation paradigm.Existing subspace learning multi-view clustering methods have achieved significant improvements,but are still confronted with challenges when tackling complex multi-view clustering tasks for the following reasons: the methods singly explore the multi-view information of the data,and ignore the close connection between subspace learning and spectral clustering;the construction of affinity matrix and the prediction of clustering indicator matrix are separate;most methods directly generate subspace representation on each original view,which insufficiently exploits the complementary information of the multi-view data.These issues limit the performance of existing methods in practical multi-view clustering applications.This paper analyses the issues in detail under image clustering scenarios,the main work is as follows:Firstly,in order to solve the problem of singly exploring multi-view information of the data and ignoring the close relationship between subspace learning and spectral clustering,this paper proposes a multi-view subspace clustering algorithm based on diverse representation and common representation(DCMSC).The algorithm can learn the diverse representation of views and the common representation of views from multi-view data and optimize the affinity matrix which not only considers the close relationship between subspace learning and spectral clustering,but also guarantees the clustering results consistent for different views.Then,according to diverse and common representation,the affinity matrix can be constructed,and clustering results can be obtained by putting affinity matrix into spectral clustering algorithm.Secondly,to deal with the problem of separating the construction of affinity matrix and the prediction of clustering indicator matrix,this paper proposes a multi-view subspace clustering algorithm based on subspace representation and indicator matrix(SIMSC).The algorithm can directly learn the multi-view subspace representation,continuous clustering indicator matrix and discrete clustering indicator matrix from the multi-view data,then the three terms with iterative optimization can make the subspace representation better to characterize the subspace structure of the multi-view data.Then,the multi-view subspace representation is used to construct the affinity matrix,and the clustering results can be obtained with spectral clustering algorithm according the affinity matrix.Thirdly,to address the problem of generating subspace representation directly on each original view,which insufficiently exploits the complementary information of the multi-view data,this paper proposes a multi-view subspace clustering algorithm based on latent similarity representation and indicator matrix(LSIMSC).This algorithm explores the similarity of data in latent subspace,and uses clustering indicator matrix to optimize latent similarity representation.Then the ability of latent similarity representation to describe multi-view data can be improved,finally the clustering results can be obtained by using latent similarity representation.In order to verify the feasibility of proposed algorithms,this paper performs experiments on the popular datasets: Olivetti Research Laboratory(ORL),Yale,and Yale-B in the field of multi-view subspace clustering.Experimental results show that the three proposed algorithms proposed in this paper are beneficial for the improvement of clustering performance.