Robust graph contrastive learning with multi-hop views for node classification

Contrastive learning has been proven successful in graph self-supervised learning by addressing label scarcity in real-world applications. Most existing methods fail to effectively incorporate multi-hop information into the augmented views but only focus on augmenting the raw input data, which incurs massive computational and memory costs. To address such limitations, we propose a novel approach, Multi-hop Views Graph Contrastive Learning (MHVGCL), that enhances the node classification performance for graphs. Specifically, in contrast to existing methods that generate multiple augmented heads via neural networks, our approach generates these heads by exploiting multi-hop information, which is obtained iteratively from a single output head. This technique can extract more comprehensive structural information without destroying the graph topology. We further devise a multi-hop contrastive loss function to maximize agreements among the multi-hop views of the same node, while minimizing them among different nodes. This design contributes to more robust representation learning that keeps structural attributes invariant under different augmented views. Numerical experimental results on a variety of benchmarks demonstrate the significant superiority of our approach over other advanced methods, by learning more discriminative node representations even with extremely limited labels.