Reproduce “ZeroG: Investigating Cross-dataset Zero-shot Transferability in Graphs” on OpenResearch

About this paper

Yuhan Li, Peisong Wang, Zhixun Li, Jeffrey Xu Yu, Jia Li

ZeroG introduces a framework for cross-dataset zero-shot node classification in graphs, leveraging large language models and parameter-efficient fine-tuning to generalize to unseen graphs with distinct label spaces. The approach achieves performance comparable to semi-supervised methods on datasets like Pubmed while significantly outperforming existing cross-dataset transfer baselines.

Key insights

• Pre-trained language models effectively create a unified semantic space for diverse graph data, addressing challenges like dimension misalignment and mismatched label spaces.
• The integration of semantically-enhanced prompting nodes and neighborhood aggregation is crucial for injecting generalizable structural and contextual information during graph pre-training.
• Parameter-efficient fine-tuning, such as LoRA, is essential for adapting large language models to graph tasks while preserving their zero-shot generalization capabilities and preventing overfitting to source datasets.

Summary by alphaXiv.