Reproduce “Self-Supervised Graph Structure Refinement for Graph Neural Networks” on OpenResearch

About this paper

Jianan Zhao, Qianlong Wen, Mingxuan Ju, Chuxu Zhang, Yanfang Ye

Researchers at the University of Notre Dame and Brandeis University developed Graph Structure Refinement (GSR), a pretrain-finetune framework that uses multi-view self-supervised learning to refine graph structures for Graph Neural Networks. This approach demonstrated superior scalability, being the only GSL method to run on ogbn-arxiv (169K nodes) without out-of-memory errors, and consistently achieved higher node classification accuracy on six benchmark datasets.

Key insights

• Decoupling graph structure learning into a self-supervised pre-training phase and a fine-tuning phase effectively resolves the goal inconsistency and scalability limitations of joint GSL frameworks.
• Multi-view contrastive learning, integrating both intra-view and inter-view objectives, enables robust learning of the underlying graph generation mechanism by leveraging complementary information across different node property views.
• Knowledge migration through pre-trained GNN weight initialization, combined with a statically refined graph structure, significantly boosts the performance and efficiency of GNNs on downstream tasks.

Summary by alphaXiv.