A Multi-Embedding Fusion Network for attributed graph clustering

Attributed graph clustering, which aims to learn embedding representation and divides nodes into different groups, has attracted increasing attention in recent years. Existing investigations have demonstrated that graph attention network (GAT) exploiting graph structure and node attributes for clustering can yield remarkable performance. However, existing GAT-based algorithms usually use adjacency matrix or feature matrix directly, neglecting the processing of noise within the feature matrix. Furthermore, some methods fail to effectively fuse different levels of embedding information for the specific clustering task. To address these deficiencies, we propose a multi-embedding fusion network for attributed graph clustering (MEFGC for short) in this paper. Specifically, in our model, a novel Laplacian filter is first designed to alleviate high-frequency noise. Secondly, we design a multi-embedding fusion module, which includes an improved auto-encoder and graph attention network, to obtain superior node embedding representation. Finally, a reliable target distribution generation method is designed, utilizing a joint supervision strategy combining self-supervision and mutual supervision to optimize the node embedding. Extensive experiments on four benchmark datasets demonstrate that the proposed MEFGC achieves state-of-the-art results in clustering tasks.