Research On Structured Representation Learning For Multi-view Clustering

In recent years,multi-view Clustering(MVC)has attracted more and more attention because it can reduce the annotation cost of massive data.Multi-view clustering aims to make use of the consistency and complementarity among views to reasonably fuse the information of multiple views and obtain better clustering performance.Most of the existing multi-view clustering methods consider the views of all samples to be complete,but in real applications,the collected data may be incomplete.Because traditional multi-view Clustering methods cannot achieve ideal performance in incomplete multi-view data,Incomplete multi-view Clustering(IMVC)becomes a challenging problem.This paper studies multi-view clustering and missing multi-view clustering respectively,and the specific research contents are as follows:(1)For multi-view clustering,most of the existing works consider using complementary information between views to improve the clustering performance,while ignoring the structural information between samples,which is of great significance for mining the correlation between samples.In this paper,a structured latent representation learning method is proposed for multi-view clustering,which considers the complementary information between views and the structural information between samples.Specifically,considering the complementarity of views,we first learn a common latent representation for all views.Then,in order to mine the structural information between samples,the knearest neighbor graph is constructed based on the common latent representation,and the graph convolution network is used to further strengthen the interaction between samples,so as to learn a structured latent representation for clustering.A large number of experiments on five real data sets demonstrate the effectiveness of the proposed method.(2)For incomplete multi-view clustering,most of the existing works only consider the information of complete views,but rarely considers the structural information between samples and the structural information of incomplete views.In this paper,we propose an integrated heterogeneous graph attention network for incomplete multi-view clustering,which not only considers the complementary information of complete views,but also the structural information between samples and the structural information of incomplete views.Specifically,we first learn a common latent representation for the complete views,then construct a knearest neighbor graph based on the common latent representation,and construct a view-specific incomplete pattern graph for each view.Then an integrated heterogeneous graph is obtained by integrating the nearest neighbor graph and the incomplete pattern graph of different views,which represents the relationship between samples under different missing patterns.Then we use the graph attention mechanism to further enhance the interaction between samples and learn a set of structured latent representations.Finally,the consistency of probability distribution is embedded into the network for clustering.Experimenting on five incomplete multiview datasets shows that the proposed method is effective and robust to view-missing.