Abstract
Clustering holds profound significance in data mining. In recent years, graphconvolutional network (GCN) has emerged as a powerful tool for deep clustering,integrating both graph structural information and node attributes. However,most existing methods ignore the higher-order structural information of thegraph. Evidently, nodes within the same cluster can establish distantconnections. Besides, recent deep clustering methods usually apply aself-supervised module to monitor the training process of their model, focusingsolely on node attributes without paying attention to graph structure. In thispaper, we propose a novel graph clustering network to make full use of graphstructural information. To capture the higher-order structural information, wedesign a graph mutual infomax module, effectively maximizing mutual informationbetween graph-level and node-level representations, and employ a trinaryself-supervised module that includes modularity as a structural constraint. Ourproposed model outperforms many state-of-the-art methods on various datasets,demonstrating its superiority.