Contrastive Learning with Cluster-Preserving Augmentation for Attributed Graph Clustering

Abstract

Graph contrastive learning has attracted considerable attention and made remarkable progress in node representation learning and clustering for attributed graphs. However, existing contrastive-based clustering methods separate the processes of node representation learning and graph clustering into two stages, making it difficult to ensure good clustering. Therefore, it remains a challenge to design an effective contrastive learning method that jointly optimizes node representations and graph clustering. Moreover, existing random augmentation strategies to generate contrastive views may destroy the original topological structures of clusters in graphs. So it is crucial to construct an augmented graph that preserves the cluster structure of a given graph while benefitting graph clustering. To address these problems, we propose a contrastive learning method with cluster-preserving augmentation for attributed graph clustering. Specifically, we construct a contrasting view based on the generated k NN graph and edge betweenness centrality to preserve the inherent cluster structure of a graph. Then, a multilevel contrastive mechanism is proposed to maximize the agreement between node representations in multiple latent spaces. Finally, the objective of node representation learning is jointly optimized with the self-supervised clustering objective to obtain cluster distributions and discriminative node representations simultaneously. Extensive experiments on seven widely used real-world graphs demonstrate that the proposed model consistently outperforms existing state-of-the-art methods on clustering tasks.