Research On Attributed Graph Clustering Based On Graph Representation Learning

Network applications,such as social networks,citation networks,and protein interaction networks,have emerged increasingly and have atracted much atention in the last decade.Unlike traditional two-dimensional data,network data is naturally represented as a graph.In addition to the pairwise structural relationship between nodes in the network,each node has its own attributes.We call this type of data attributed graphs,and effectively analyzing attribute graphs provides users with a deeper understanding of what’s behind the data,and has shown great popularity in addressing community detection,enterprise social network user grouping,and more.Therefore,mining attribute graphs has important academic significance and application prospects.Attributed graph clustering aims to divide nodes into different clusters by making full use of node attributes and graph structure.The main work includes graph representation learning based on graph node features,graph structure,and graph convolutional network(GCN).Models based on graph node features and graph structure focus more on one aspect of node features and topology,but lack the fusion of node features and mining topology.Models based on GCN update node representations by combining adjacent node features to build graph embeddings,The existing methods also have the following problems:(1)Relying on the original graph structure,the graph embedding is obtained through the similarity inferred by the topology structure,and the similarity between node features is ignored in the training process.(2)Graph embedding learning mainly uses non-cluster-driven losses and cannot learn friendly clustering features.(1)Two models are proposed,a deep graph embedding network model based on feature and structure similarity(FSDGEN),and attributed graph clustering model based on fused multimodal autoencoder(MAEGC)to solve the attributed graph clustering problem.(2)In order to effectively capture the underlying information of the feature space,this thesis adopts a feature graph construction method,which obtains the similarity between features according to KNN and constructs the feature graph.Through the designed feature graph autoencoder,it is convenient for node features to propagate in the feature space to extract feature information.In order to effectively utilize the high-order neighbor informati on,this thesis adopts the structure graph attention network to capture the structure information.(3)The variational graph autoencoder is combined with attribute graph clustering for the first time,which reduces the strong dependence on node features in the clustering process to a certain extent.Meanwhile,in order to learn friendly clustering representations,Adopt dual self-supervised clustering and self-supervised clustering to guide graph representation learning.(4)Extensive experiments demonstrate the superior performance of this new method over state-of-the-art methods.