Dual Self-Supervised Deep Graph Clustering

Contrastive deep graph clustering has garnered significant attention owing to its robust representation capacity and superior clustering performance. However, existing contrastive deep graph clustering methods typically treat different views of the same node as positive samples and different views of different nodes as negative samples, neglecting the importance of various nodes and the structural information of the graph. To alleviate this issue, we propose a novel method calledDualSelf-SupervisedDeepGraphClustering (DSDGC). Specifically, to generate different views, we employ unshared Multi-Layer Perceptrons (MLPs) to project the pre-processed graph into various subspaces as different views. Then, we design the first self-supervised learning module,i.e.,a weighted contrastive module, by introducing the cross-view similarity matrix as a weight matrix for positive/negative sample selection to bring similar nodes closer while pushing dissimilar nodes apart. Moreover, we align the semantic labels with high-confidence clustering pseudo-labels, forming the second self-supervised module, to further guarantee the clustering reliability. Extensive experiments on six benchmark datasets demonstrate that the proposed DSDGC method outperforms existing state-of-the-art methods.