Research On Subspace Clustering Algorithm Based On Self-expression

The complexity of its high-dimensional data is reflected in: high dimensionality,complex structure,and a large number of irrelevant attributes.Traditional clustering method based on sample spacing to construct clusters,such as K-means?spectral clustering and so on,cannot directly process high-dimensional data.Now,with the mature application of subspace clustering technology to high-dimensional data processing,the majority of scientific research scholars have developed multi-view subspace clustering and self-expression subspace clustering algorithms.The work done in this paper for the above two types of algorithms is as follows:Existing self-representation subspace clustering algorithms lack the processing of related data,resulting in too dense affinity matrix and the algorithm contains wrong information,or too sparse to cause the algorithm to lose relevant information.At the same time,it shows the lack of consistency between the affinity matrix and the segmentation matrix,resulting in suboptimal results obtained by the algorithm.Therefore,this paper proposed the weighted structured correlation low-rank subspace clustering(SCLR).This algorithm solves the problem of too dense affinity matrix under low-rank characteristics by introducing weighted correlation norms;Meanwhile,this model adds structured norms,integrates affinity matrix learning and spectral clustering segmentation into a framework,forcing the two matrices to be related and consistent.The results obtained on the Extended Yale B and MNIST datasets show that SCLR performs better than other algorithms.Most current multi-view subspace clustering algorithms do not pay attention to the block diagonal structure of the matrix,and this structure has a non-negligible effect on the subspace segmentation effect.Although the affinity matrix of most algorithms can obtain the block diagonal structure,but when the data contains noise,etc.,the block diagonal structure will be destroyed.Therefore,this paper block diagonal representation regularized multi-view subspace clustering algorithm with low-rank constraint(BDR-MSCLT).This model combines the low-rank tensor-constrained multi-view subspace clustering model and soft block diagonal norm.On the one hand,the soft block diagonal constraint enhances the block diagonal structure of the affinity matrix,making the structure more robust.On the other hand,the algorithm integrates the affinity matrix of different views into tensor,capturing the high-order correlation of data.Finally,the low-rank characteristics of tensors preserve the global structure of the data and reduce the redundant information between the representation matrices.The results under several classical experimental datasets show that the BDR-MSCLT algorithm is more robust and various clustering indicators are better.