Experimental evaluation demonstrates zero-shot graph learning across diverse relational networks, highlighting effective generalization without task-specific training.
Key Points
OpenGraph achieves robust zero-shot graph learning across unseen relational networks, successfully overcoming topological distribution shifts between training and test sets.
The architecture pairs a unified graph tokenizer with a scalable graph transformer, encoding global dependencies and using language models for structural data augmentation.
The framework demonstrates effective cross-domain transfer for node classification and link prediction, suggesting viable pathways toward universal graph foundation models.