Coordinated Representation Based Multi-view Subspace Clustering

With the rapid development of the information technology and the widespread use of data collection sensors,data produced and collected in daily life and manufacturing fields are becoming increasingly complex.Apart from high-dimensional features,the form of data is increasingly diverse,such as multi-descriptiveness,multi-modality,and multi-source,etc.Multi-view high-dimensional data is one of the most common data forms in the era of big data,and corresponding clustering problem is a research hotspot in the fields of pattern recognition,machine learning,etc.Multi-view data provides comprehensive description from different perspectives,and there are complex semantic correlations between different views.Thus,learning the coordinated representations of multi-view data to mine their complementary information and exploring the structural consistency and correlation among cross-view representations are the key problems of accurate clustering of complex multi-view data.It is also of great significance to enrich and develop the theory and method of multi-view learning.In recent years,the theories of subspace representation learning and spectral clustering have provided a new way to solve the problems of efficient representation and clustering of multi-view high-dimensional data,and there are some rich research results.However,the existing multi-view clustering methods are hard to maintain the essential characteristics of multi-view data(representation complementarity and consistency of clustering structure),and cannot effectively model the correlation between cross-view representations,seriously affecting the performance of multi-view data clustering applications.Based on the above analyses and considerations,focusing on representation learning and cross-view semantic correlation modeling in the multi-view clustering tasks,based on the compatibility and complementarity,this thesis studies the coordinated subspace representation learning method of multi-view data and establishes different levels of cross-view representation correlation modeling and mining strategies.Then several kind of improved multi-view subspace clustering methods are proposed.At the same time,for the clustering task where only single-view data is available due to special scenarios or extreme view missing of multi-view data,focusing on the problems of limited description information and insufficient utilization of intrinsic attribute information of single-view data,a multi-attribute subspace clustering method with intrinsic attributes enhancement is developed.The main innovations and contributions of this dissertation are summarized as follows:First,in view of the lack of effective perception of global structure information in the process of multi-view subspace representation learning for existing multi-view subspace clustering models,the global block diagonal representation coupled multi-view subspace clustering method is designed,which achieves effective balance between cross-view shared information and view-specific information.This method combines local view-specific representation learning and global consistent representation fitting in a unified framework,in which global representation with clear clustering structure is used to respond to view-specific representation learning and reinforce each other.Therefore,it could aggregate the specific information of multi-view representations to capture complementary global information and enhance consistent description.The experimental results of multi-view clustering tasks on multiple real datasets demonstrate its effectiveness.Second,for the problems that weak complementarity of multi-view representations caused by mining cross-view consistency information with element consistency separation strategy from representation level and high computational complexity caused by low-rank constraint,the rank consistency induced multi-view subspace clustering model is developed.To facilitate a practical model,all multi-view subspace representations are parameterized through the matrix tri-factorization along with orthogonal constraints.Thus,representation complementary and structure consistent multi-view low-rank subspace representations are coordinately learned,ensuring the complementary attribute of multi-view representations and effectively capturing the structural consistency correlation across multi-views.On this basis,a mathematical model that is easy to optimize and corresponding optimization algorithm with lower computational complexity are developed.Experimental results demonstrate that this method significantly improves clustering performance and greatly reduces running time.Third,low-rank tensor could effectively fuse complementary information and model higher-order correlation across views in the unified tensor space.The existing methods usually adopt the convex tensor nuclear norm(TNN)to characterize low-rank structure and thus ignore the physical difference of singular values,resulting in multi-view tensor representation with suboptimal view-specific representations and redundant cross-view correlation.This thesis studies logarithmic schatten-p function regularized tensor nuclear norm,which implicitly considers the difference of singular values.Theoretical analysis shows that the Tensor Logarithmic Schatten-p Nuclear Norm(TLS_pNN)has stronger rank function approximation ability.On this basis,the tensor Logarithmic Schatten-p nuclear norm based multi-view subspace clustering model is proposed,in which TLS_pNN is used to constrain the multi-view subspace representation tensor.Thus,the learned tensor representation with compact and discriminative low-rank structure will well explore the complementary information and accurately characterize the high-order correlation among multi-views,reducing intra-view representation redundancy and inter-view correlation redundancy.Experimental results on some challenge and complex datasets show that the proposed method has significant advantages in clustering performance and robustness.Finally,for the clustering task where only single-view data with limited description information is available due to special scenarios or extreme view missing of multi-view data,the existing methods only learn a single subspace representation from whole features of original data,failing to fully depict and understand data.They ignores the inherent multi-attribute information of data,to address this issue,a multi-attribute subspace clustering method with data intrinsic attributes enhancement is developed.This method explores the multi-attribute information of the data and learns complementary multi-attribute subspace representations based on complete multi-attribute latent features,and then capture high-order correlation among multi-attribute representations in a compact and coupled low-rank tensor space.Thus,attribute-enhanced complementary information is used to comprehensively describe the data,which expands the width and depth of data understanding.The experimental results show that the proposed method has obvious advantages over single-view clustering methods;On some datasets with complex attributes,its clustering performance is comparable to or even better than the multi-view methods,which fully validates the effectiveness of data’s intrinsic attributes enhancement.