Research On Multi-view Clustering Method Based On Subspace Learning

With the development of multimedia technology,data is exploding.Many real-world data can be described by different views or representations in various multimedia applications.Therefore,multi-view learning researches are of great significance.Multi-view clustering has drawn much attention recently.The purpose of subspace clustering is to discover the potential subspace structure of data,and calculate the corresponding affine matrix,and finally implement spectral clustering algorithm on the affine matrix,so as to find the subspace where the data belongs and get the clustering results.Therefore,in this paper,two solutions based on the subspace model are proposed for the case of intact view and incomplete view,and they are validated effectively.The main contributions of this paper are as follows:(1)In the case of complete view,a subspace-based model is proposed in this paper.Subspace clustering methods aim to find the underlying subspaces of data.However,for multi-view subspace clustering,the shared subspace information has not been fully utilized.Therefore,in our method,a shared subspace is learned to preserve the effective consensus information of all views.Then,an affinity matrix with adaptive neighbors is learned to assign the most suitable cluster to each data point.Finally,experiments on several benchmark datasets demonstrate that the proposed algorithm outperforms other state-of-the-art algorithms.(2)In the case of incomplete view,another subspace-based model is proposed in this paper.Traditional multi-view clustering methods assume that each view has complete data.However,missing data are more common.To address the problem,our method seeks the latent space and performs data reconstruction simultaneously to learn the subspace representation.In addition,we enforce the subspace representation to be non-negative,yielding a weight interpretation among different data.Finally,experiments on one synthetic dataset and several benchmark datasets validate the effectiveness of the proposed subspace-based method.(3)For the solutions of the above two problems,this paper proposes an alternating minimization method based on an effective alternating search strategy.In addition,for both solutions,we provide the corresponding convergence analysis.