In the information era,the data is growing explosively,most of which are unlabeled data.Therefore,clustering analysis techniques for unsupervised learning of data are becoming more and more important.At present,machine learning algorithms are mainly run on data represented by a single view,which is actually contrary to the way that humans analyze problems comprehensively from multiple angles.The opposite of single-view clustering is multi-view clustering,which runs clustering algorithms on data sets from multiple feature sources,and can take advantages of the complementarity between views to improve the clustering effect on the original single view.Subspace learning and spectral clustering are the most widely used algorithms in multi-view clustering at present.However,the existing multi-view spectral clustering algorithms based on subspace learning are generally based on low-rank or sparse priors,and ignore the relationship between the original data samples,so that the captured clustering structure is not accurate enough.In addition,in traditional spectral clustering algorithms,the independent solving steps of clustering labels often lead to suboptimal solutions.To solve the above problems,in order to future improve the multi-view clustering performance,this paper focuses on clustering structure mining and clustering label learning,and proposes two multi-view clustering models.The main work is as follows:The first model is multi-view subspace clustering based on consistent similarity learning(CSL).The model is based on subspace learning.It learns low-dimensional subspace representations that are more conducive to clustering tasks from high-dimensional data through selfexpression models with the guidance of the similarity graph of the original data.Then it learns the consistent similarity matrix from the new characteristic representation of each subspace under the condition of constrained Laplacian rank.The learned consistent similarity matrix has a more ideal block diagonal structure,so that subsequent spectral clustering graph cuts can obtain more accurate clustering results.On the basis of the first model,this paper further proposes an integrated multi-view clustering algorithm based on subspace learning and spectral rotation.The Laplacian matrix is merged to more sensitively capture the discrepancies between different views.The spectral rotation theory is utilized to take the place of k-means step of spectral clustering,resulting in directly obtaining the discrete cluster label within an unified clustering framework.The variables are optimized in an unified process thus can be fed back and guided by each other.Hence,the obtained clustering result is closer to the groud truth cluster label.In conclusion,the main work of this paper is:(1)Analyze the key points of the multi-view clustering problem,and propose corresponding improvements to the current shortcomings of the multi-view spectral clustering algorithm based on subspace learning.Two models CSL and UMSSC are proposed through reasonable multi-view fusion strategies with the focus on cluster structure mining and cluster label learning,respectively;(2)The optimization algorithms corresponding to the two models are proposed,the solving steps are described in detail,and the convergence and complexity of the algorithms are analyzed,with the experimental evaluation conducted.(3)Model CSL and UMSSC were compared with 12 the state-of-the-art methods on 6 real world benchmark datasets,and the parameters of the two models were analyzed respectively.In addition,ablation experiments were carried out on the innovation points of this paper.Experimental results prove that the two models in this paper can effectively improve the clustering effect and are suitable for a variety of multi-view data scenarios. |