Attention-Based Graph Clustering with Enhanced Delivery Operator

Recently, graph clustering methods that integrate structured information into deep clustering have become one of the important research directions. However, current methods still lack an efficient and flexible fusion strategy when combining auto encoders and Graph Convolutional Network (GCN) to learn feature representation. Accordingly, we propose a network named Structural Attention-based Graph Clustering Network (SAGC). Firstly, we use the attention coefficient as weighted values to dynamically allocate the contribution of node attribute features and structural features. Secondly, an enhanced delivery operator is designed to efficiently convey the node features learned by the autoencoder. Lastly, a dual self-supervised mechanism is employed to unify the optimization of the two types of node features, thereby enhancing the stability of the model. Experiments on various types of datasets validate that our method outperforms state-of-the-art methods.