Graph Contrastive Learning via Weisfeiler-Leman Dual-View Sampling

Abstract: Graph contrastive learning (GCL) approaches have gained momentum over the past few years. By augmenting the original graph data, the common GCL pipeline learns from such multiple contrastive graph views in a self-supervised manner, tackling critical issues in the literature, such as node label scarcity. To obtain contrastive views, most GCL techniques heavily rely on feature-space similarity measures. We consider this as a limiting factor in GCL, since it implies that node features are (in general) informative and closely aligned with the graph topology, an assumption that does not hold, for instance, in the case of heterophilic graphs. In this work, we propose to address the problem by coupling the usual feature-space similarities with structure-based measures, which we propose to implement through the Weisfeiler-Leman (WL) family of algorithms. Our framework, dubbed WLGCL, introduces a dual-view sampling strategy that works on features- and WL-level to construct more reliable contrastive pairs. WLGCL integrates a multi-positive and hard-negative contrastive loss to ensure the alignment-uniformity trade-off without modifying the design of the graph encoder. Extensive experiments on six benchmark datasets and against seven state-of-the-art baselines demonstrate the efficacy of WLGCL, where additional empirical evaluations justify the adoption of our architectural choices for the model.

Submission Type: Long submission (more than 12 pages of main content)

Assigned Action Editor: ~Samuel_Vaiter1

Submission Number: 8755