Research On Multi-view Clustering Based On Sparse Representation

Multi-view clustering is a popular research direction in the domain of machine learning.Its major task is to fuse data of different views,obtain potential patterns of the data and then divide the data into various categories according to the data properties of multi-view data.At present,researchers have proposed many methods to study multi-view clustering.Among them,the multi-view clustering methods based on representation learning attract more attention.The core idea of the methods is to learn a unified representation from multi-view data in subspace for clustering.Thus,the key of the multi-view clustering methods based on representation learning is how to learn a highly discriminant representation from multi-view data.In recent years,sparse representation methods have been widely used in many fields due to the existence of the prior knowledge of sparsity in effective information of data.Some researchers have applied sparse representation to study multi-view clustering and achieved good clustering results,but there is still room for further improvement.This thesis give a description of the research background and significance about multi-view clustering and sparse representation algorithm firstly,then analyzes the research status of sparse representation learning models and multi-view clustering in detail,summarizes the problems occurring in the existing multi-view clustering method,and designs the corresponding solutions and methods.The major contributions of this thesis are as follows(1)In view of the methods of existing multi-view clustering only takes into account the properties of multi-view data such as consistency and complementarity,or the redundancy of information within each view,this thesis has designed a sparse representation multi-view learning method with l1 norm regularization constraint.The method employs a partial sharing policy to make use of consistency and complementarity for acquiring the potential patterns of multi-view data.And then it adopts the l1 norm regularization constraint to the information redundancy within every view to get a high quality representation.Besides,the graph regularization constraint is introduced to preserve the intrinsic manifold structure of data and further improve the clustering effect.At the same time,an efficient optimization algorithm is designed to speed up the training of the multi-view learning model.Experimentally,on six real multi-view datasets,the proposed method has some advantages over the existing single-view/multi-view clustering methods.(2)On account of the existing multi-view methods based on sparse representation has failed to make full use of the properties of multi-view data,and ignored the local structure between the samples,this thesis has developed a sparse representation multi-view learning method with locality-constrained linear coding.At the same time,the graph regularization constraint is adopted to preserve the potential manifold structure of data and further improve the clustering effect.In addition,according to the proposed multi-view learning model,an effective optimization algorithm is designed to train the model.Experiments show that the proposed method is superior to existing multi-view clustering methods on four general multi-view datasets.