- Rating
- 4
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 3
Read review
Paper summary
The paper proposes GraphVec, a language-model-free framework for cross-domain graph-level representation learning. The method has several key components, such as multi-scale global relational graphs, density-maximization alignment, and hierarchical reference distribution encoding. The paper provides an end-to-end generalization bound connecting source contrastive pretraining to few-shot classification on an unseen target domain. GraphVec is evaluated by few-shot graph classification on diverse graph datasets from biology, social network, and computer vision. It also reports graph clustering, node classification, ablations, robustness, scalability, and hyperparameter sensitivity. GraphVec substantially outperforms recent graph pretraining, prompt-learning, and GFM baselines in several cross-domain settings.
Critical strengths
-
The key techniques of GraphVec, including multi-scale global relational graphs, density-maximization alignment, and hierarchical reference distribution encoding are novel, well motivated, and clearly verified.
-
GraphVec achieves large gains over strong baselines in cross-domain few-shot graph classification, e.g., average 66.14 vs. 58.63 for GCN and 58.40 for BRIDGE. The paper also evaluates unsupervised pretraining, graph clustering, node classification, ablations, robustness to noisy graphs, and scalability. This breadth makes the empirical contribution convincing.
-
The convergence analysis for density-maximization alignment and the end-to-end generalization bound are nontrivial and go beyond typical empirical GFM papers. The theory connects contrastive pretraining, alignment, and target few-shot classification in a principled way, which materially strengthens the paper.
Critical weaknesses
The global multi-graph construction requires kernel matrices and SVD over all nodes in a dataset. Although Nyström approximation is mentioned, the reported peak RAM usage is already high for a moderate dataset, and the method may struggle with very large graphs.
Questions for authors
-
Since density-maximization alignment only matches spectral means and does not guarantee removal of higher-order or task-relevant distributional differences, does this undermine robustness and generalization in harder cross-domain settings?
-
The LLM-based GFMs perform poorly in several comparisons. Given that node features are numerical rather than natural-language attributes, how fair is the text conversion protocol? Could the weak LLM performance be an artifact of representation collapse caused by the conversion?
-
How does GraphVec scale to much larger graph datasets? The reported peak RAM is high? Can the method handle million-node graphs or thousands of domains without prohibitive memory?
Discussion reminder
Dear Reviewers,thank you for your time and thoughtful evaluation of the Paper. The authors have now submitted their rebuttal addressing the questions and concerns raised in your initial reviews. Please carefully review the authors' responses alongside the other reviews, and participate in the discussion to share your updated thoughts and comments. To help us reach a well-informed meta-review decision, please let us know if the rebuttal adequately resolves your concerns and whether you plan to adjust your final rating. Please submit your post-rebuttal comments and updated evaluation on the review portal. Thank you again for your valuable service and contribution to the review process. AC