Multi-view Clustering Algorithm Based On Subspace Learning

In recent years,with the development of information technology,the means by which human beings acquire information become more diversified,resulting in a large amount of multi-view data in many scientific and industrial fields.How to extract useful information from multi-view data has become a hot topic in machine learning.Multi-view clustering aims at dividing multi-view data into different clusters according to the potential structural information of multi-view data,and has achieved satisfactory clustering results,so it has gradually received extensive attention from the academic and industrial circles.Multi-view clustering based on subspace learning models the original high-dimensional feature space by multiple linear subspaces,and effectively solves the clustering problem of high-dimensional multi-view data.While existing multi-view clustering algorithms based on subspace learning have achieved good performance,but they still face great challenges when dealing with multiple view data in a complicated environment,since multi-view data often have different characteristics,dimensions and structure,and view consistency,complementarity and correlation between various clustering information.Mining clustering information hidden in multi-view data is a key and technical difficulty in multi-view learning,and it is also an open problem that is widely concerned.Therefore,this paper is committed to designing an effective multi-view clustering algorithm to fully mine the useful information of multi-view data,so as to effectively improve the clustering performance.The main research works and contributions are summarized as follows:(1)To solve the "dimensional disaster" of high-dimensional multi-view data and the problem that existing algorithms fail to fully mine the consistency and diversity of multi-view data,this paper proposes a novel multi-view clustering method,named Diversity and Consistency Embedding Learning(DCEL).DCEL learns a better affinity matrix in a learned latent embedding space while considering diversity and consistency of multi-view data simultaneously.Specifically,by a projection technology,DCEL projects raw data into a latent embedding space to obtain the low dimensional latent representation of multi-view data.Then,utilizing the self-representation nature of data,DCEL learns a shared consistent representation from multiple views,as well as a diverse representation of each view,to learn an optimal relational matrix in an embedded space.On this basis,an optimization algorithm based on alternating direction multiplier method(ADMM)is proposed to solve the objective function of DCEL.Finally,experimental results on five common data sets show that DCEL performs better than some of the latest multi-view clustering methods.(2)To solve the problem that existing algorithms only use single-layer mapping to mine the structural information of data,this paper proposes a novel multi-view clustering method,namely Multi-view Clustering via Deep Level Semantics Exploiting(DLSE).DLSE mines the deep semantic information of data and learns the final cluster indicator matrix under a unified framework.Specifically,a new paradigm,Deep matrix factorization(DMF),is designed to mine the hierarchical semantics of the original data through layer by layer,and forces the samples from same cluster in multiple views to get closer together in low dimensional space.In addition,in order to preserve the local geometry of the original data in the low-dimensional representations obtained in DMF,this paper introduces local preserving regularization to guide DMF.Meanwhile,a new spectral rotation fusion paradigm is designed to directly learn the final clustering indicator matrix during decomposition.Finally,a large number of experimental data prove the superiority of DLSE algorithm.(3)To solve the problem that existing algorithms ignore the high-order information of multi-view data,this paper proposes a novel multi-view clustering method,namely Multi-view Clustering with Dual Tensors(MCDT).MCDT makes use of both the high-order correlation within views(that is,the correlation between different samples in the same view)and the highorder correlation between views(that is,the correlation between the same sample in different views).Specifically,MCDT first learns a set of view-specific affinity matrices by subspace self-representation learning for each view.Next,these affinity matrices are then stacked into a third-order tensor.Then,the tensor low-rank constraint is applied to mine higher-order correlations within the view.At the same time,MCDT rotates this third-order tensor to mine high-order correlations between views,to more fully exploit high-order information hidden in multiple views.A large number of experiments on multiple benchmark datasets show that MCDT achieves better clustering performance than the existing advanced multi-view clustering algorithms.To sum up,this paper focuses on the multi-view clustering task,fully mines the clustering information hidden in multi-view data,and realizes the mining of diversity and consistency,deep semantic information and high-order correlation,and systematically proposes three multi-view clustering methods based on subspace learning.