SEGCN: Structural Enhancement Graph Clustering Network

Deep graph clustering, which reveals the intrinsic structure and underlying relationship of graph node data, has become a highly concerning and challenging research task for graph-structured data. In recent years, existing graph clustering methods have gotten better performance without human guidance by combining auto-encoder and graph convolution networks. However, the existing methods exist problems: 1) the aggregation of noise information in the graph structure information leads to poor model learning effect; 2) the methods focus on learning the local structure information of the graph, and fail to learn the global structure information of the graph, resulting in incomplete learning of the graph. To overcome the shortcomings, we propose a $\mathbf{S}$tructural $\mathbf{E}$nhancement $\mathbf{G}$raph $\mathbf{C}$lustering $\mathbf{N}$etwork($\mathbf{SEGCN}$) to learn the graph topology information hidden in the dependencies of nodes attribute and the global structure information. Specifically, SEGCN adopts the method of calculating correlation to extract the graph topology information hidden in the dependencies of node attributes and design a dynamic fusion strategy to enrich the original graph structure information. Meanwhile, SEGCN enriches the embedding information extracted from the local structure by fusing information on the global structure based on the global structure dynamic fusion method. Moreover, to cluster the subsequent, SEGCN applies the pseudo-cluster labels learned by a dual self-supervised strategy to guide the learning of the node attribute representation and optimize the weight matrix of networks. Extensive experimental results on four benchmark datasets indicate that SEGCN achieves excellent performance on all datasets and accomplishes an accuracy of 71.56% and NMI of 44.50% on the CITE dataset.