A Flexible, Equivariant Framework for Subgraph GNNs via Graph Products and Graph Coarsening

Subgraph GNNs enhance message-passing GNNs expressivity by representing graphs as sets of subgraphs, demonstrating impressive performance across various tasks. However, their scalability is hindered by the need to process large numbers of subgraphs. While previous approaches attempted to generate smaller subsets of subgraphs through random or learnable sampling, these methods often yielded suboptimal selections or were limited to small subset sizes, ultimately compromising their effectiveness. This paper introduces a new Subgraph GNN framework to address these issues. Our approach diverges from most previous methods by associating subgraphs with node clusters rather than with individual nodes. We show that the resulting collection of subgraphs can be viewed as the product of coarsened and original graphs, unveiling a new connectivity structure on which we perform generalized message passing. Crucially, controlling the coarsening function enables meaningful selection of any number of subgraphs. In addition, we reveal novel permutation symmetries in the resulting node feature tensor, characterize associated linear equivariant layers, and integrate them into our Subgraph GNN. We also introduce novel node marking strategies and provide a theoretical analysis of their expressive power and other key aspects of our approach. Extensive experiments on multiple graph learning benchmarks demonstrate that our method is significantly more flexible than previous approaches, as it can seamlessly handle any number of subgraphs, while consistently outperforming baseline approaches. Our code is available at https://github.com/BarSGuy/Efficient-Subgraph-GNNs.

Paper

References (41)

Scroll for more · 29 remaining

Similar papers

Peer review

Reviewer DujM7/10 · confidence 2/52024-07-06

Summary

This paper proposes CS-GNN, a subgraph GNN approach utilizing graph coarsening and graph Cartesian products. This novel and flexible subgraph GNN can effectively generate and process any desired bag size. The paper also discovers new permutation symmetries in the produced node feature tensor during generalized message passing. Theoretical and empirical analysis validate the efficacy of the proposed method.

Strengths

1. The topic of subgraph GNNs is interesting. 2. The paper is generally well-written with clarity. 3. The proposed method is novel, with theoretical underpinnings. 4. The empirical results seem promising. 5. The implementation code is provided for reproducing the results.

Weaknesses

1. The size of the coarsened graph, though sparse, is still very large with $2^n$ nodes, which might result in high computation complexity for large input graphs. 2. Eq. (3) indicates that the input graph is unweighted and edge weights might not be considered.

Questions

1. The size of the coarsened graph, though sparse, is still very large with $2^n$ nodes, which might result in high computation complexity for large input graphs. Do you have an estimate of the complexity and runtime of the proposed model compared with others? 2. Eq. (3) indicates that the input graph is unweighted and edge weights might not be considered. What if the input graph is weighted, for example, with edge weights 1,2,3, and 4.

Rating

7

Confidence

2

Soundness

3

Presentation

3

Contribution

3

Limitations

Yes, the authors have adequately addressed the limitations and, if applicable, potential negative societal impact of their work.

Reviewer h2GK6/10 · confidence 4/52024-07-12

Summary

The authors introduce a novel graph learning method that leverages message passing over product graphs. Specifically, message passing is performed over both the input graph and its product with a coarsened version of itself, which can be derived through techniques such as node clustering. Additionally, the authors propose a concept called symmetry-based updates, where message passing is enhanced by incorporating a graph representing a linear equivariant layer over the product graphs. A basis for such layers is also developed as part of their contribution. Experimental results demonstrate that this method performs comparably to state-of-the-art approaches.

Strengths

**S1 Product construction** The main conceptual contribution of the paper is the use of product graph construction to allow more flexibility and control over certain types of subgraph GNNs. The product construction itself is quite natural, although it introduces some overhead depending on the chosen instantiation. **S2 Equivariant basis** The key theoretical contribution is the identification of the bases of equivariant layers for such product graphs. This allows for a principled introduction of message passing using symmetries in the product graph. **S3 Competitive architecture** Experiments validate the proposed methods, showing that they are competitive with state-of-the-art methods.

Weaknesses

**W1 Unclear key definition** The key idea underlying the paper is commendable, although not extremely innovative. However, one crucial component is the graph coarsening introduced in Section 4.1. I found the definition on page 4 confusing despite multiple readings. What is defined there is an exponentially large fixed graph, with all subsets as nodes and adjacencies defined in equation (3). However, you mention before and after this that various coarsenings can be used, such as spectral clustering. How do these fit into the fixed definition given here? The presentation could be clearer. **W2 Weak theoretical analysis** The significance of Section 5, where the authors use their model for analyzing marking strategies, is not strongly demonstrated, at least based on the main paper. It primarily serves to justify the marking strategy in equation (10) but leaves a more detailed understanding of this strategy open. The section would benefit from a clearer presentation. **W3 Unclear presentation** Certain parts, as noted already in **W1** and **W2**, are not clearly described. Another example is Section 4.2.3, which is too succinct, and the visualization of $A_{equiv}$​ is rather cryptic. You mention a characterization and updates, but it is not clear what is meant here. A clear description of the instantiation of the framework used would be helpful. Similarly, the section starting from line 357 about “more than a coarsening function” is unclear, including the “informal” proposition statement. **W4 Unclear connection with OSANs** In the related work section, the authors contrast their proposal with OSANs [30], stating that OSANs use ordered tuples whereas the current paper uses sets of indices. However, unordered subgraphs play a key role in the OSAN paper as well since they showed that they can encompass most subgraph formalisms. A more detailed comparison would benefit the paper. **W5 Missing experimental details** - For the peptides experiments, you write “by using effective subsets of subgraphs.” What do you mean by this? What is this effective subset? How did you choose it? - Which version of OSAN do you compare with in the experiments (Table 4)? - Why do you use comparisons with different methods across the experiments (Table 2 vs Table 4)? **W6 Maximally expressiveness: clarification needed** At several points in the paper, you mention maximally expressive subgraph GNNs. It is unclear what is meant by this and this should be detailed, regardless of whether it was shown in [39]. As far as I know, general subgraph GNN do not have a maximally expressive instantiation, so I assume certain restrictions on the class of subgraph GNNs are in place. It would be beneficial to make these explicit to better understand the limitations of the proposed methods. **W7 Minor comments** *All figures (especially the font size used in the text in figures) in the paper are too small. Please enlarge them.* - L.82: What is $\mathcal{X}$? - L218: The choice of $S_n$​ as a symbol for the symmetric group is not optimal, as subsets are denoted by $S_1$, $S_2$, etc. - L225: You mention “extension to multiple features is straightforward” and refer to [22]. A bit more explanation would be helpful.

Questions

I would appreciate it if the authors could respond to the weak points **W1**, **W4**, **W5**, and explain what they want to convey with the section on “more than a coarsening function" in Section 5.

Rating

6

Confidence

4

Soundness

3

Presentation

1

Contribution

2

Limitations

This has been addressed in a satisfactory way by the authors.

Reviewer hoHT2/10 · confidence 3/52024-07-12

Summary

The paper is very well written. Existing subgraph GNN methods are mentioned. Then a novel subgraph GNN framework is formulated. The authors discuss the equivariance properties of this new subgraph GNN formulation followed by an experimental evaluation of the proposed method.

Strengths

* Strong mathematical formulation of the developed methods is provided. * The ideas are communicated well. One can see that the authors put a lot of time and effort into this work. * A comprehensive appendix is provided explaining the proposed approach in even more detail.

Weaknesses

### The use of few and small datasets If the goal of this work is to keep computational complexity low while improving the model quality, strong experimental results are expected. This puts into question the developed theory and methods in the paper. I suggest the authors to find a scenario where their model can achieve SOTA results to justify their ideas. I would recommend including more datasets such as PCQM4Mv2 [1] to the experimental evaluation. ### Limited baselines A simple comparison against a recent linear runtime complexity graph transformer alternative [2] shows that the experimental results are weak. A select few datasets are chosen and the achieved results are well below the current SOTA on them. ### Weak motivation If subgraph GNNs address the issue of limited expressiveness of existing work as stated in the introduction of the paper, then it makes one question why are the experimental results not good. [1] https://ogb.stanford.edu/docs/lsc/pcqm4mv2/ [2] https://openreview.net/pdf?id=duLr8BIzro

Questions

1. Is there a case where the computational gains leads to a SOTA results on a large scale dataset? 2. Could the experimental results improved if the proposed approach was combined with an approach similar to [2]. Improving the expressibility of the proposed model even further by combining with other classes of GNNs to get SOTA results would significantly improve the chances of the paper. [2] https://openreview.net/pdf?id=duLr8BIzro

Rating

2

Confidence

3

Soundness

1

Presentation

4

Contribution

2

Limitations

The authors admit their proposed method has a high computational complexity. Thus naturally one would expect even higher model quality.

Authorsrebuttal2024-08-12

Relating to our response

Dear Reviewer hoHT, We appreciate your thoughtful review and hope that we have adequately addressed your concerns. We made an effort to address your feedback during the rebuttal period. Specifically, we shared new results on new experiments: * We conducted new experiments using a large-scale dataset OGBG-MOLPCBA (437K graphs). Additionally, we referred you to another large-scale experiment we already included in the original submission using ZINC FULL (250K graphs). Both experiments show solid results for our method compared to SoTA. * We successfully incorporated components from [2] to enhance our architecture. These new results can be found in our original response and PDF. We would be grateful if you could review these additions and consider raising your score if they satisfactorily address your concerns. Best regards, The Authors

Reviewer hoHT2024-08-12

I would like to thank the authors for their rebuttal. I am comparing Table 2, Peptides-func and peptides-struct results to [2]. The reported SOTA results in [2] for these datasets are 0.6975 and 0.2464, both of which significantly are better than what is being reported in Table 2 in this work. molhiv results are subpar as well (79.44 in this paper vs 80.84 reported in for CIN method and 79.80 for GECO both in [2]). Essentially there is no dataset that justifies the added complexity of the proposed methods in this work. To me, it looks like the authors are working on making an approach faster that does not seem to bring any clear benefits. > The goal of this work is to reduce the time and space complexity of Subgraph GNNs. As a result, some model degradation is expected. Subgraph GNNs are provably expressive models with theoretical guarantees [3, 4, 5], which may offer certain advantages over MPNNs and graph transformers. If subgraph GNNs have high computational complexity but don't work well, then this undermines the impact of the work in this paper. Since my score is borderline, I would like to instead change my score and go for a recommendation of rejection. If the authors can motivate subgraph GNNs better by showing that they work well over a class of datasets in a future revision of the paper, that would significantly boost the chances of acceptance of their work. This is due to the fact that datasets with a lot of graphs already are made of small graphs, and non-subgraph classes of methods already work well and are computationally efficient. [2] https://openreview.net/pdf?id=duLr8BIzro

Authorsrebuttal2024-08-13

Concerns Regarding Unjustified Score Reduction

Dear Reviewer, We appreciate your time and effort in evaluating our work. However, we respectfully disagree with the recent score reduction and would like to raise some concerns: 1. Changing the score so abruptly just before the discussion deadline hinders our ability to respond and compromises the fairness of the rebuttal and discussion process. 2. Our rebuttal comprehensively addressed all your initial concerns, providing clarifications and additional positive experimental results as per your request. Your response does not acknowledge our inputs or new results. Instead, it lowers the score from 4 to 2 based on the same initial comments that originally received a score of 4. This discrepancy is difficult to reconcile. 3. Your current score of 2 contrasts sharply with both your initial assessment and the positive evaluations from other reviewers (6,7). 4. We are concerned that our prompt for a response may have indirectly led to this unjustified score reduction. 5. We would like to emphasize a key contribution that may have been overlooked: the introduction of new symmetries and the characterization of accompanying equivariant layers. For a conference of this tier, we expect reviews to be consistent, thorough, and responsive to rebuttals. The dramatic score change without new concerns or acknowledgment of our responses does not align with these expectations. We have also contacted the AC and SAC regarding this matter. We respectfully request that you reassess your recent score reduction.

Reviewer h2GK2024-08-09

Thanks for the rebuttal

Dear authors, thank you for your detailed rebuttal and thoughtful responses to my concerns. Incorporating those comments will definitely improve the paper.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC