| With the advancement of time and technology,the amount and type of data that needs to be processed in life is growing dramatically.For example,images shared on websites often have corresponding text tags and descriptions,and a single image is described by different types of features,which is called multi-view data.It is becoming increasingly difficult to extract useful knowledge from the huge amount of data.Clustering analysis is a very important data mining method,but most of the traditional clustering algorithms are shallow single-view clustering algorithms.Deep learning has shown its powerful representation learning capability in various fields,which can capture the nonlinear features in data.Graph neural networks are the application of deep learning on graph data.In this paper,we study multi-view clustering algorithms based on graph neural networks,and propose attention fusion based graph convolutional networks for multi-view clustering(AFGCN)and contrastive multi-view graph clustering network(CMGCN).For attribute multi-view graph data,AFGCN uses graph convolutional networks to construct multi-view graph self-encoders,with each view corresponding to a graph self-encoder.The representation of the data is learned by the multi-view graph self-encoder.Since different perspectives have different importance in clustering,an attention fusion network is designed.It obtains better clustering results by focusing on the representation of important perspectives.In addition,a weight loss function is designed based on the prior knowledge to embed the prior information into the model.Finally,the representation learning and clustering tasks of attribute multiview graphs are unified in a single framework,and both tasks benefit from each other by iteratively performing representation learning and clustering.Experiments show excellent results of this method on multiple datasets.AFGCN deals with attribute multi-view graph data,but there are some data in real life that are multi-attribute multi-view graph data.For this reason,CMGCN is proposed to solve the clustering problem of multi-attribute multi-view graph data.First,CMGCN uses data augmentation to construct a new perspective on the data.Secondly,a shared coefficient matrix is obtained at the self-expression layer using multi-view subspace clustering.To prevent learning to degenerate solutions,contrast learning is used to constrain the coefficient matrix in addition to using Fparametrics as coefficient loss on the coefficient matrix.Specifically,contrast loss is used to make nodes similar to their neighbors in the graph structure and dissimilar to other nodes.Finally,spectral clustering is performed on the coefficient matrix to obtain clustering results.Experiments show that the method has good performance on several datasets. |
