Graph contrastive clustering network with randomized Schur complement augmentation

Contrastive graph node clustering necessitates multiple augmented views for contrastive learning, commonly generated by perturbing node attributes or edges. However, these methods are prone to semantic drift and may fail to preserve the original graph's topological structure, thereby undermining the reliability of contrastive learning. To mitigate these challenges, this paper introduces the Graph Contrastive Clustering Network with Randomized Schur Complement Augmentation (CRSCA). Specifically, by applying Gaussian elimination to the adjacency matrix, we generated randomized Schur complement matrices as augmented views. This method ensures that the random walk transition probabilities of nodes relative to the original graph are maintained, thereby preserving the graph's structure and topology. By integrating autoencoders and graph neural networks, CRSCA demonstrates its effectiveness through extensive experiments conducted on five datasets.