Research On Multi-View Subspace Clustering Algorithm Based On Low Rank Representation

With the massive production of multi-view data,clustering analysis of multi-view data has increasingly received widespread attention from researchers.Each view of multi-view data has different characteristics and carries different amounts of information.How much information should be used for different views in the clustering process,and how to obtain a high accuracy multi-view data fusion strategy,remains a challenge that needs to be addressed.There is correlation between multi-view data,and data from different views also contains information that other view data does not have.Therefore,when obtaining the subspace representation of multi-view data,how to preserve the diversity features between different views when obtaining subspace representations is a very important issue.In this paper,we studied multi-view subspace clustering algorithms based on these two issues.The following research results were obtained:(1)The existing multi-view subspace clustering algorithms ignored the situation where different views have different amounts of information when obtaining a common subspace,which can easily caused information loss and affect clustering results.A multi-view subspace clustering algorithm based on information entropy weighting was proposed to solve the problem of different information content in each view during multi-view subspace clustering.Firstly,a low rank representation was used to constrain the subspace learning of each view,and then a common representation was used to ensure consistency.When determining the common representation,the weight of each view was determined through information entropy,then the algorithm was optimized using Lagrange multiplier method.Finally,a large number of experiments were conducted on five data sets to verify the effectiveness of the algorithm.(2)When obtaining the subspace representation of each view,considering the inconsistency of information between different views,we use Hilbert-Schmidt Independence Criterion,HSIC as an independence metric to fully mine the diversity information in each view,and propose a multi-view subspace clustering algorithm based on HSIC.The main work includes: firstly,we used low rank representations to constrain the subspace learning of each view,then introduced symmetric constraints to perform approximately linear spatial transformations on multi-view data,and then added diversity penalty terms to obtain a subspace representation of multiple views based on HSIC.The Lagrange multiplier method was used to optimize the above algorithm.Finally,for spectral clustering of multi-view data,an affinity matrix was constructed using symmetry constraints to obtain the final results.A large number of experiments on 8 data sets had verified the effectiveness of the algorithm.