Reliable Node Similarity Matrix Guided Contrastive Graph Clustering

Abstract

Graph clustering, which involves the partitioning of nodes within a graph\ninto disjoint clusters, holds significant importance for numerous subsequent\napplications. Recently, contrastive learning, known for utilizing supervisory\ninformation, has demonstrated encouraging results in deep graph clustering.\nThis methodology facilitates the learning of favorable node representations for\nclustering by attracting positively correlated node pairs and distancing\nnegatively correlated pairs within the representation space. Nevertheless, a\nsignificant limitation of existing methods is their inadequacy in thoroughly\nexploring node-wise similarity. For instance, some hypothesize that the node\nsimilarity matrix within the representation space is identical, ignoring the\ninherent semantic relationships among nodes. Given the fundamental role of\ninstance similarity in clustering, our research investigates contrastive graph\nclustering from the perspective of the node similarity matrix. We argue that an\nideal node similarity matrix within the representation space should accurately\nreflect the inherent semantic relationships among nodes, ensuring the\npreservation of semantic similarities in the learned representations. In\nresponse to this, we introduce a new framework, Reliable Node Similarity Matrix\nGuided Contrastive Graph Clustering (NS4GC), which estimates an approximately\nideal node similarity matrix within the representation space to guide\nrepresentation learning. Our method introduces node-neighbor alignment and\nsemantic-aware sparsification, ensuring the node similarity matrix is both\naccurate and efficiently sparse. Comprehensive experiments conducted on $8$\nreal-world datasets affirm the efficacy of learning the node similarity matrix\nand the superior performance of NS4GC.\n