Research On Multi-view Clustering Method Based On Consensus Matrix Learning

In recent years,with the rapid development of information collection and processing technology,multi-view data,which are collected from different domains or represented by diverse features,can be seen everywhere in practical applications.The rapid growth of multi-view data brings both opportunities and challenges to the field of data mining.Due to the lack of class label information in multi-view data,people cannot accurately obtain the desired data.Therefore,it is particularly important to cluster multi-view data.Multi-view clustering usually aims at maximizing the consistency information of data on different feature views and accomplishes the clustering task by learning a consensus matrix from multi-view data that can best characterize the common underlying structure shared by different views.However,with the complexity of multi-view data attributes,noise,high-dimensional and nonlinear structure of multi-view data will affect the quality of the learned consensus matrix.This dissertation focuses on the factors that affect the quality of the learned consensus matrix during multi-view clustering,such as large noise,inability to capture the internal structure information of the data,high dimensionality,lack of rich fused information,nonlinear structure and a series of factors,and carries out research on multi-view clustering method based on consensus matrix learning.The main research contents include the following four aspects:(1)Aiming at the problems that the consensus affinity graph matrix learned by the existing multi-view clustering methods based on graph model can hardly capture the intrinsic structure of the data,cannot distinguish the importance of different views,and does not have an explicit cluster structure,thus it is not qualified for the task of clustering,this dissertation proposed a multi-view clustering based on consensus affinity graph matrix learning via graph fusion.This method captures the intrinsic structure of data in each single view by constructing a structure graph fusion framework,and when learning the consensus affinity graph matrix,a weighted fusion strategy is introduced,which considers the discrimination of different views and incorporates the complementary information and intrinsic structure information of important views.At the same time,based on the inter-view consistency principle,the consensus similarity information that is unanimously admitted by all views is learned.A block diagonal regularizer constraint is introduced to the consensus affinity graph matrix so that it has an explicit cluster structure.Experimental results show that by capturing intra-view graph structure,weighted fusing the inter-view complementary information,learning the inter-view consistency similarity information,and using structured block diagonal representation,the learned consensus affinity graph matrix can achieve better clustering performance.(2)Aiming at the problem that the existing high-dimensional multi-view data cannot preserve the main discriminative features and structural features of the original data during dimensionality reduction and cannot resist the negative impact of noise,outliers,and corruption,which leads to the problem of learning an unreliable and inaccurate affinity matrix in the embedded space,thus degenerating the clustering performance,this dissertation proposed a multi-view clustering based on consensus affinity matrix learning in embedding space.In this method,a double-relaxed energy-preserving embedding model is proposed,which uses two energy-preserving embedding routes to preserve more discriminative information of the original data in the embedding space.Locality preserving projection and low rank preserving embedding are used to preserve the local and global structural information of the original data in the embedding space.The energy loss error and the data reconstruction error in the embedding space are learned to enhance the robustness against noise,outliers,and corruption.Experimental results show that the consensus affinity matrix learned by this method can obtain better clustering performance,and the information loss is small in the embedding process.(3)Aiming at the problem that most existing multi-view clustering methods only exploit the consistency information or complementarity information among multiple feature views in the original space or latent embedding space,which limits their ability to learn an informative consensus representation,this dissertation proposed a multi-view clustering based on enhanced consensus representation matrix learning.The proposed method employs adaptive neighbor graph learning to construct an initial affinity matrix for each view as input.A latent representation correlation preserving model is proposed.By establishing the correlation between the shared latent representation of the initial affinity matrices in the latent embedding space and the latent representation of the consensus affinity matrix in the latent embedding space,the improved consensus representation matrix can not only incorporate the consistency information in the original space but also reconstruct the complementary information in the latent embedding space.Experimental results show that the learned consensus representation matrix can provide useful information for clustering and achieve better clustering performance.(4)Aiming at the problem that the existing kernel multi-view clustering methods only focus on kernel function learning and ignore the graph structure learning in kernel space,a kernel multi-view clustering based on consensus affinity graph matrix learning is proposed.This method constructs candidate kernel graphs that can preserve the local and global structure of the data in the kernel space.A consensus affinity graph is learned jointly from all the candidate kernel graphs and the latent representations of these candidate kernel graphs,which effectively reduces the negative effect of noise and avoids the problem of fusing too much redundant information that may be brought by the previous method of directly fusing all the candidate kernel graphs.Experimental results show that by using kernel function to explore the possible nonlinear structure in the data and learning the graph structure in the kernel space,the learned consensus affinity graph matrix can obtain better clustering performance.In summary,the four proposed methods verify that a clustering-friendly consensus matrix can be separately learned by mining the intrinsic structure of the multi-view data,seeking an efficient energy and structure preserving embedding approach for the original high-dimensional multi-view data,simultaneously exploiting the available information in both the original space and the embedding space,and mining the nonlinear structure of multi-view data.Thus,the clustering performances of multi-view data can be improved.