Research And Application On Multi-view Clustering Algorithm Based On Subspace Learning

In the background of rapid development,the number and dimensionality of data are increasing day by day,which can be described by different views,and the traditional singleview clustering research can no longer meet this trend.Compared with single-view clustering,multi-view clustering has become a hot research topic nowadays by mining the potential connections between views and classifying the data effectively.(1)Diversity-induced Multi-view Clustering in Latent Embedded Space is proposed.Traditional multi-view clustering algorithms consider that multi-view data partitioning should be in its original feature space,and the merit of its clustering results relies heavily and implicitly on the quality of the original feature presentation.In addition,it is important to use the view to recover potential space and diversity information.To address the above problems,this paper proposes Diversity-induced Multi-view Clustering in Latent Embedded Space algorithm,which consists of four parts.First,the concept of latent embedded space model is introduced for latent embedded learning,assuming that the latent embedded space can represent each view.Second,global similarity learning of the potential embedding space ensures the accuracy of the global structure.Third,the projection matrix specific to the view is constrained by an empirical Hilbert Schmidt Independence Criterion to obtain diversity information between views.Fourth,the clustering indicator is learned so that the number of connected components of the consistency affinity matrix is equal to the number of clustering clusters.An alternating optimization scheme is also designed to integrate latent embedded learning,diversity learning,global similarity learning and cluster indicator learning in a same framework.Comparative experimental results on six publicly available datasets show that the algorithm has some advantages.(2)Smoothness Regularized Multi-view Clustering with Local Structure Learning is proposed.The existing multi-view clustering assumes that all views enjoy the same coefficient matrix,this approach ignores the differences between views,in practice,and the coefficient matrix should have the same clustering features instead of being identical.Secondly,there are heterogeneous features between views in multi-view data,and these heterogeneous features represent the local structure of different views.Smoothness Regularized Multi-view Clustering with Local Structure Learning to address the above problem.The algorithm consists of three parts.First,a factorization structure with orthogonality constraints and low-rank constraints is introduced to ensure view consistency structure.Second,increase the local structure learning to enhance the local structure characteristics of different views and increase the local structure differences brought about by the differences between different views.Third,smoothness regularized learning is used to smooth the fluctuations caused by the local structure and make the algorithm more stable.An Augmented Lagrange Multiplier method is also designed for alternating iterative optimization for solving the objective function,and the consistency affinity matrix obtained from the iterations is put into the spectral clustering framework to calculate the final clustering results.The results of the comparison experiments on six benchmark datasets show the effectiveness of the algorithm.(3)Application of multi-view clustering algorithm to Water Grades classification.The Water Grades reflect the water quality of an area region,and has good social significance for pollution prevention and control and pollution management.With the increase in the number of Water Grades indicators,how to read meaningful information from these evaluation indicators has some research value.The two multi-view clustering algorithms proposed in this paper are applied to Water Grades classification to obtain more accurate Water Grades.Compared with several algorithms,the experimental results show that the two multi-view clustering algorithms proposed in this paper have some practical value.