Multi-view Clustering Algorithm Based On Kernel Method And Subspace Learning

In recent years,data information has shown an explosive growth trend,and multi-view data composed by different sources or with multiple feature descriptions are valued.Multi-view clustering methods can characterize data from different perspectives and effectively reveal the internal structure information of data,therefore,multi-view clustering research gradually becomes a hot topic nowadays.With the depth of practical research,there are many problems to be solved.First,the consistency and complementarity of multi-view data often cannot be reconciled,and the complementarity property is more easily ignored in multi-view clustering.Secondly,when multi-view clustering deals with nonlinear data,such data cannot be well divided in the construction of the adjacency matrix,which affects the quality of the adjacency matrix and also fails to capture the nonlinear information in it.Finally,if the views with different quality have the same importance in the view fusion stage,it will affect the effect of information fusion.Based on the above,this dissertation proposes a multi-view clustering study based on kernel method and latent representation.The main studies are as follows:1)The existing multi-view clustering model learns the linear representation of data based on Euclidean distance measure in the process of constructing the adjacency matrix,while the processing of nonlinear data cannot achieve good data separation and fully explore the nonlinear characteristics of data,which affects the construction of adjacency matrix.The kernel method is applied to cluster analysis because it can effectively model the nonlinear relationship between data points.By projecting the data from the current low-dimensional original feature space to the new high-dimensional space through the kernel method,the separability of the data set is improved,and then the adjacency matrix of each view is obtained to fully explore the nonlinear information of the data.Based on this,a multi-view spectral clustering algorithm based on the kernel method(MSCKM)is proposed.MSCKM uses kernel techniques to fully exploit the nonlinear relationships between the data when constructing the adjacency matrix,and solves for each view to obtain the adjacency matrix and the corresponding spectral embedding matrix.And then,uniformly solving the classification indicator matrix.Finally,multiple kernel functions are combined to learn the best combination of kernel functions to achieve the fusion of view information.The results of the comparison experiments verify the effectiveness of the method.2)The common multi-view subspace clustering tends to focus on the consistency information between views,but ignores the complementarity information between views.It also does not take into account the rich heterogeneous information of views and the prevalent noise of views.The noise can negatively affect the clustering results.Complementarity information and view noise are common and important influencing factors of multi-view clustering,They are often not focused and optimized at the same time.Based on this,a self-weighted multi-view subspace clustering based on latent low-rank sparse constrain(SMSC-LLSC)is proposed.Specifically,the potential representation of the multi-view subspace is learned to comprehensively obtain the complementary information that each view contains that is not contained in other views,and the potential representation matrix has the characteristics of low rank sparse,thus making the potential representation of the subspace more accurate;in the process of constructing the shared representation matrix,adaptive weights are assigned to each view,the weights measure the degree of contribution of each view to the clustering effect,and the weights are determined by the inverse relationship of the distance between the representation matrices of the different views and the uniform representation matrix.The experimental results verify the effectiveness of the method.In this dissertation,we perform multi-view clustering by considering the complementary properties and the widespread nonlinearity of multi-view data,as well as the differences between views and the uneven quality of views.Separable operations on nonlinear data and acquisition of nonlinear information are achieved through kernel methods;the potential representation of jointly learned views and view adaptive weights are used to better capture the complementary information between views and reduce the negative impact of low-quality views on the clustering effect.The experimental results and comparative analysis on several multi-view datasets,as well as the corresponding statistical hypothesis testing analysis,show the effectiveness of the two algorithms proposed in this dissertation.