Shared-Attribute Multi-Graph Clustering with Global Self-Attention

Jianpeng Chen, Zhimeng Yang, Jingyu Pu, Yazhou Ren, Xiaorong Pu, Li Gao, Lifang He

Abstract: Recently, multi-view attributed graph clustering has attracted lots of attention with the explosion of graph-structured data. Existing methods are primarily designed for the form in which every graph has its attributes. We argue that a more natural form of multi-view attributed graph data contains shared node attributes and multiple graphs, which we called “multi-graph”. When simply applying existing methods to multi-graph clustering, the information of shared attributes is not well exploited to eliminate the large variances among different graphs. Therefore, we propose a Shared-Attribute Multi-Graph Clustering with global self-attention (SAMGC) method for multi-graph clustering. The main ideas of SAMGC are: 1) Global self-attention is proposed to construct the supplementary graph from shared attributes for each graph. 2) Layer attention is proposed to meet the requirements for different layers in different graphs. 3) A novel self-supervised weighting strategy is proposed to de-emphasize unimportant graphs. Our experiments on four benchmark datasets show the superiority of SAMGC over 14 SOTA methods. The source code is available at https://github.com/cjpcool/SAMGC.