Research On The Ensemble Method Of Deep Subspace Clustering

Due to the strong representational learning ability of deep network and its ability to handle non-linear features in data well,the fusion of deep network and clustering model has become the trend of unsupervised learning.In this context,the authors propose a method of Deep Subspace Clustering Networks(DSC-Nets)based on selfexpression peculiarity,which embedding self-expression peculiarity into deep Autoencoder Networks.Combining the advantages of deep learning and subspace clustering,the results show superior clustering performance.However,DSC-Nets is sensitive to the structure of autoencoder networks,and the performance of the subspace it learns is restricted by the number of layers and other factors.For this reason,this thesis proposes an ensemble learning method to fuse multiple subspaces generated by deep subspace clustering networks to improve the final clustering performance,and has achieved remarkable results.The main innovations of this thesis are as follows:Firstly,the Sequential Ensemble of Deep Subspace Clustering(SeqEn-DSC)is proposed.SeqEn-DSC introduced randomness by using alternate iteration training and random cover,and sequentially generated a group of diversified self-expression coefficient matrix,and then obtained the similarity matrix and integrated it sequentially.The sequential ensemble method does not need to store a large amount of base clustering,after which a new similarity matrix is directly added to the integration,thus keeping a small memory consumption.Experiments on the benchmark datasets show that the sequential ensemble method has better clustering performance than the parallel ensemble and other single subspace clustering methods,and can achieve robust results for different network structures.Furthermore,On the basis of SeqEn-DSC framework,this thesis proposes a Selfensemble of Deep Subspace Clustering(SelEn-DSC).Self-ensemble learning is used to guide deep subspace clustering to generate more diversified self-expression coefficient matrices.Compared with the sequential forward learning framework of SeqEn-DSC,SelEn-DSC adds a self-ensemble module to update the similarity matrix of the current integration.At the same time,the ensemble similarity matrix reversely directs the self-expression module learning.Through the supervision of the selfensemble module and the constraint of subspace clustering on the self-expression layer,the accuracy and diversity of the new similarity matrix continuously generated in the iterative process are guaranteed.Experiments on benchmark datasets show that the proposed framework can effectively improve the performance of DSC-Nets and Deep Low-Rank Subspace Clustering(DLRSC),and is also superior to the existing Deep Subspace Clustering methods.