Partial Multi-view Subspace Clustering Based On Deep Learning

Multi-view subspace clustering assumes that the data of each view is complete when dealing with the clustering problem of multi-view data with high dimension.However,data is always missing in practical problems,which leads to the problem of partial multi-view subspace clustering.Considering that the unsupervised network structure in deep learning has strong expression ability,and can extract high-level feature representation,this paper mainly combines deep learning with subspace learning to explore the partial multi-view subspace clustering method.Most of the existing multi-view subspace clustering methods only learn a common clustering structure,and can not make full use of the existing data information to infer the missing data.Based on Star GAN and subspace learning,paper proposes a partial multi-view subspace clustering algorithm SSPMVC.SSPMVC designs a generation model based on Star GAN,make full use of the existing data information to generate all the missing data of view,and only need to train a generator.The generation model captures the global structure of view data diversity and consistency.Then,based on the diversity of multi-view data,a new multiview subspace clustering algorithm is proposed.SSPMVC trains the generation model and clustering model together,and optimizes the two model alternately.Experiments on four real data show that SSPMVC algorithm has excellent clustering performance.Most of the existing multi-view clustering methods based on subspace learning adopt a two-stage strategy,and only consider the consistency or diversity of multi-view data.Therefore,this paper proposes an end-to-end partial multi-view subspace clustering method DFLSL?PMVC based on subspace learning and deep network.DFLSL?PMVC designs two subnetworks: diversity network and consistency network.DFLSL?PMVC learns the diversity relationship and consistency relationship of multi-view data,which is an end-to-end unified framework to realize feature learning,similarity matrix learning and clustering process.DFLSL?PMVC achieves the alignment from the multiple representation of data to the consistency of clustering results,and better mines the potential subspace structure of view data.The paper also gives the training strategy for the complex network structure.Experimental results on four real data show that,compared with the advanced partial multi-view clustering algorithm,DFLSL?PMVC still achieves good clustering performance.