Unifying Attribute and Structure Preservation for Enhanced Graph Contrastive Learning

The graph contrastive learning (GCL) has garnered significant interest due to its strong capability to capture both graph structure and node attribute information through self-supervised learning. However, current GCL frameworks primarily construct final contrastive views based on local structure view, neglecting the valuable complementary information provided by the attribute view. In this article, we aim to effectively incorporate the attribute view into GCL while leveraging multiscale structure views. We identify that directly contrasting the attribute view with the local structure view results in impaired performance, primarily due to the excessively low level of mutual information (MI) between these two contrastive views. To overcome this inherent limitation, we propose a novel Attribute and Structure-Preserving GCL framework, named attribute and structure-preserving graph contrastive learning (ASP). ASP adopts an innovative contrastive view generation process that aggregates different graph views as the final contrastive view. The framework has two main modules: the attribute-preserving contrastive learning module and the structure-preserving contrastive learning module. These modules capture attribute and long-range global structure information of the input graphs. We further extend ASP to ASP-adaptive which can flexibly generate contrastive views with adaptive aggregation mechanisms. Extensive experiments on real-world graph benchmarks demonstrate the superiority of ASP and ASP-adaptive over several representative baselines on both node classification and link prediction tasks. The source code is available at: https://github.com/JialuChenChina/ASP-adaptive.