Reversible and irreversible bracket-based dynamics for deep graph neural networks

Recent works have shown that physics-inspired architectures allow the training of deep graph neural networks (GNNs) without oversmoothing. The role of these physics is unclear, however, with successful examples of both reversible (e.g., Hamiltonian) and irreversible (e.g., diffusion) phenomena producing comparable results despite diametrically opposed mechanisms, and further complications arising due to empirical departures from mathematical theory. This work presents a series of novel GNN architectures based upon structure-preserving bracket-based dynamical systems, which are provably guaranteed to either conserve energy or generate positive dissipation with increasing depth. It is shown that the theoretically principled framework employed here allows for inherently explainable constructions, which contextualize departures from theory in current architectures and better elucidate the roles of reversibility and irreversibility in network performance.

Paper

Similar papers

Peer review

Reviewer dWFZ5/10 · confidence 2/52023-07-06

Summary

This paper provides a unified framework inspired the bracket-based dynamical system to analysis the oversmoothing problem in GNN. The past work may leverage the opposite physics concept such as the reversible processes, irreversible process and therefore it is not clear how such concept help to design GNNs, while the framework in this paper gives a deeper understanding. Leveraging the data-driven exterior calculus, this paper constructs four novel architectures which span the both reversibility and irreversibility spectrum using geometric brackets as a means of parameterizing dynamics abstractly without empirically assuming a physical model. Interestingly, it can reinterpretate the message-passing and GAT as the fluxes and conservation balances of physics simulator. It also generalize the attention mechanism which extends the nodal feature to higher order cliques and provide a unified evaluation of dissipation. Emperically, the author compare the architectures derived from this framework with classical GNN,e.g., GAT, GDE over several benchmarks.

Strengths

1. This paper is well-written and offers valuable insights into the construction of graph neural networks. As far as I am aware, the idea presented in this paper is innovative. However, since I lack a background in physics, I would appreciate hearing the suggestions of other reviewers who are familiar with bracket-based dynamical systems. 2. The experimental result in damped double pendulum and MuJoCO dynamics looks good. 3. The author made the code available, promoting reproducibility.

Weaknesses

1. This paper may be a little hard to understand for the reader without physics background. 2. It seems that the algorithm proposed by the author does not outperform the baselines in node classification problem.

Questions

I am curious about the precise computational complexity of the architectures proposed by the author, specifically in the context of N nodes or N neighbors. It would be beneficial to have concrete values rather than just the order of complexity. Additionally, it would be valuable to understand the advantages of these architectures in comparison to traditional graph neural networks (GNNs). Regarding the node classification experiment, the algorithm's performance is generally comparable to the baseline, although it does appear weaker in some cases. However, it showcases excellent performance in physical systems like the double pendulum and MuJoCo dynamics. Could this be attributed to the notion that physics-inspired frameworks are inherently more suitable for describing and modeling physical systems rather than social networks?

Rating

5: Borderline accept: Technically solid paper where reasons to accept outweigh reasons to reject, e.g., limited evaluation. Please use sparingly.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

Yes

Reviewer 7QNL6/10 · confidence 2/52023-07-06

Summary

This work provides a comprehensive overview of graph attention networks (GATs), shedding light on their fundamental concepts and principles. Additionally, the authors introduce a set of novel GNN architectures that leverage structure-preserving bracket-based dynamical systems. By incorporating these systems into GNNs, the proposed architectures aim to enhance the modeling capabilities and performance of graph-based machine learning tasks. Overall, the paper presents both a broad perspective on GATs and innovative approaches to further advance the field.

Strengths

They give a detailed theoretical analysis of graph attention networks with exterior calculus techniques. They also provide a generalized attention mechanism and propose several novel structure-preserving GNNs with satisfying performance.

Weaknesses

The table writing is not clear. Sometimes orange result is better than blue result while sometime blue is better.

Questions

I'd like to know whether your analysis can explain why "Gradient“ model performs worse on heterophily graphs like Computer and Photo. Which physics system is suitable for heterophily graphs and which is suitable for homophily graphs?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

They have already listed their limitations.

Reviewer y9dF6/10 · confidence 3/52023-07-06

Summary

The work proposes structure-preserving bracket-based dynamical systems to learn physical systems using GNNs. The authors proposed four formulations depending on completeness and character (conservative or dissipative). The models are demonstrated to be effective in physical system and node classification tasks.

Strengths

* The method is constructed based on solid mathematics, and incorporating attention as a learnable inner product seems quite natural. * The authors suggested various formalisms depending on completeness and character under the notion of the abstract bracket formulations, which offers a general viewpoint for physical systems. * The method is quite adaptive because it is demonstrated to be effective in various tasks, including physical system and node classification.

Weaknesses

* It is unclear to the reviewer where is the authors' contribution and where it is from prior works, particularly the mathematical part. The presentation could be more clear to distinguish existing work and the contribution of the work. * The authors claim, "We use these architectures to systematically evaluate the role of (ir)reversibility in the performance of deep GNNs." However, the reviewer could not catch "the role of (ir)reversibility" from the experiments. The authors could elaborate on this point in the experiments part. * Despite the broad range of experiments considered, the performance of the models is moderate. Therefore, it is unclear in which condition the proposed models are effective (see the following weakness for more details). * In Section 5.1, the Hamiltonian model works better than the gradient model on the dissipative phenomena. The author could explain why the conservative works better here as it could be key to understanding "the role of (ir)reversibility." * In addition, there are no experiments where the gradient model works best. With these results, it is unclear when the gradient model is effective. The authors could add experiments where the gradient works fine, e.g., incomplete and totally dissipative system. * In Section 5.3, the authors reported that the proposed model consumed much memory. In this case, computation time for prediction could also be reported because there are typically tradeoffs between speed and accuracy. Minor points: * Pointers to "Table 3" (49, 124, and 159) might be "Table 1."

Questions

* Does the method work well for PDEs, e.g., heat and Navier–Stokes equations?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

3: You are fairly confident in your assessment. It is possible that you did not understand some parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

3 good

Contribution

3 good

Limitations

The authors adequately addressed the limitations.

Reviewer Ui2h6/10 · confidence 2/52023-07-25

Summary

This paper proposes a bracket-based dynamical system framework to design structure-preserving graph neural networks (GNNs). Specifically, the authors leverage four formalisms: Hamiltonian, Gradient, Double-Bracket, and Metriplectic, that model physical systems with different completeness and dissipation characteristics, and implement them via discretizations from exterior calculus. The authors parameterize GNNs with a chosen dynamic by specifying the bracket matrices and learning the energy/entropy functions via graph attention mechanisms. Experiments on real-world graph inference tasks show that structure-preserving GNNs can perform better than those that are not strictly structure-preserving, and the best dynamics depend on the downstream task.

Strengths

Strengths: 1. It is novel to use bracket-based parameterizations to design physics-inspired GNNs. 2. It is fruitful to analyze existing GNNs (e.g., GAT, GRAND) under the bracket-based dynamical framework.

Weaknesses

Weaknesses: 1. Since the paper is intended for general audience in graph machine learning, it can be written in a clearer manner with more motivating examples and targeting applications. For starters, include a notation section. 2. One of the key motivations of the paper is to improve performance of deep GNNs, yet all the experiments in the main manuscripts do not investigate explicitly the role of depth - are the GNNs used in these experiments actually deep? The more related experiments are described and discussed in Appendix B.3.2 as ablation study, but results (Table 8) show that for node classification task, only Gradient-based GNNs perform well when the number of layers increases, whereas other dynamics exhibit (significant) performance drop. If the paper intends to mitigate over-smoothing, the authors should provide more compelling experiments. On the other hand, if the paper simply intends to describe a physics-inspired framework to parameterize GNNs, then the authors should re-write the discussion of deep GNNs.

Questions

1. What's the implications for practitioners? What applications would we expect imposing certain bracket structure outperforms the unconstrained ones? What structural requirement one should choose for different applications (e.g., add a column in Table 1 giving example applications)? The current set of experiments reads a bit confusing to me. For example, experiment 2 considers the node classification task, which requires some sense of clustering / smoothing, but not over-smoothing. So the findings in Table 4 showing that Double Bracket performs the best aligns with our expectations? But in Table 8, the Double Bracket formalism cannot be computed for deeper architectures - Is it merely a computational constraint? More explicit remarks and discussions will be great! 2. The role of (learnable inner product) $A$: the authors connect this with graph attention mechanism; but a priori one can also choose other forms of $A$? Have the authors compare some simple baselines, say fixed $A$ or even just set $A$ as identity? This could also help address the memory issue in experiments (Sec 5.3) where the authors cannot compute $A$ (and thus $E, S$) due to the high-dimensionality of the node features? 3. The authors discussed higher-order attention mechanism based on higher-order cliques (Appendix 4). Will this be useful in practice or is this a pure theoretical discussion? If the former, can the authors provide experimental demonstration? If the latter, the authors might consider removing this from the primary contributions since this does not seem like the focus of the paper?

Rating

6: Weak Accept: Technically solid, moderate-to-high impact paper, with no major concerns with respect to evaluation, resources, reproducibility, ethical considerations.

Confidence

2: You are willing to defend your assessment, but it is quite likely that you did not understand the central parts of the submission or that you are unfamiliar with some pieces of related work. Math/other details were not carefully checked.

Soundness

3 good

Presentation

2 fair

Contribution

3 good

Limitations

This work does not have negative societal impact. The limitations are discussed in Section 6.

Reviewer Ui2h2023-08-12

Clarification of Table 9 and 10: interpretation of depth

Thank you for your detailed response! I have two follow-up questions below: Question: 1. The new Table 9 caption reads "Note that integration time can be considered as a surrogate for depth, since the temporal step-size of each network is fixed to 1". On the other hand, in the global response, the authors comment that "...but now progressively reduces the step-size of their 4th-order Runge-Kutta time integration from 1 to 1/64 ...". So Table 9 still uses a fixed step-size, while Table 10 uses a decreasing step-size? 2. If we adopt the interpretation of "integration time $\approx$ depth when step size is fixed to 1" in Table 9, then shouldn't the same analogy hold for Table 10, where the **"64" layers should effectively be much smaller**, since it should be a sum of $1 + 1/2 + 1/3 \ldots 1/64 \approx \ln 64 + \gamma \approx 4.7$ (i.e. the 64-th harmonic number)? If so, is the interpretation of "64" layer somewhat misleading, especially comparing to "standard GNNs, including GATs, are known to experience significant degradation after ~4 layers"?

Authorsrebuttal2023-08-14

Thanks for your continued interest in our paper. Please see our responses below for answers to your questions: 1. *The new Table 9 caption reads "Note that integration time can be considered as a surrogate for depth, since the temporal step-size of each network is fixed to 1". On the other hand, in the global response, the authors comment that "...but now progressively reduces the step-size of their 4th-order Runge-Kutta time integration from 1 to 1/64 ...". So Table 9 still uses a fixed step-size, while Table 10 uses a decreasing step-size?* Not exactly. The final time of integration for each bracket architecture is fixed according to the value in Table 9, and the step-size is constant for each entry in Table 10. This means that, for each column in Table 10, we choose step_size = final_time / num_layers, so that time integration is performed at variable degrees of temporal resolution to test stability with increasing depth. 2. *If we adopt the interpretation of "integration time $\approx$ depth when step size is fixed to 1" in Table 9, then shouldn't the same analogy hold for Table 10, where the "64" layers should effectively be much smaller, since it should be a sum of $1 + 1/2 + 1/3 \ldots 1/64 \approx \ln 64 + \gamma \approx 4.7$ (i.e. the 64-th harmonic number)? If so, is the interpretation of "64" layer somewhat misleading, especially comparing to "standard GNNs, including GATs, are known to experience significant degradation after ~4 layers"?* Again, there is no step-size decrease during the inference procedure, and the meaning of depth is the same in both cases. During network training, the step-size is fixed to 1. Therefore, a final time of, e.g., 11, means that (in a forward Euler scheme) 11 network compositions are performed to reach the network’s predictions, which is why we say integration time is approximately depth in Table 9. This interpretation continues in Table 10, where the step-size is changed to create a variable number of network compositions for the same final integration time: from 2 up to 64. This shows that the performance of the bracket-based architectures does not degrade with increasing depth, in contrast to more standard GNNs. Let us know if this answers your questions. If not, we are happy to continue this discussion.

Reviewer Ui2h2023-08-15

Thank you and further clarifications

Thank you for your explanation. However, in the global response, you wrote " This new study, which has been designed to supplement Table 8 in Appendix B.3.2, takes our bracket-based networks trained on CORA with Planetoid splits, but now progressively **reduces the step-size of their 4th-order Runge-Kutta time integration from 1 to 1/64** and tests their predictive accuracy on random splits." While you clarified above saying in Table 9-10, the step-size is constant. Do you refer to different "step-size" parameter in the global response versus your response here? Please clarify. Thanks!

Authorsrebuttal2023-08-15

I understand the confusion now, thanks for clarifying. We perform a refinement study, showing results taking 2 steps, 4 steps, 8 steps...64 steps. For each experiment, the timestep size is 1/numbersteps.

Reviewer Ui2h2023-08-18

Clarification of refinement study

Thank you for the clarification. Is the refinement study the same as the "depth study" in Table 10, or a separate study?

Authorsrebuttal2023-08-20

Yes, the reviewer's understanding is correct.

Reviewer dWFZ2023-08-16

Thanks for the reply

I am glad to see that the authors plan to add some more motivating examples for the reader without physics background. I have no further questions.

Reviewer y9dF2023-08-16

Thank you for the detailed responses, which addressed my concerns raised in the first review. Assuming these responses are incorporated in the revised manuscript, I have raised my scores (presentation: 2 -> 3, rating 4 -> 6). I generally appreciate the mathematical contribution and the broad range of experiments made by the work. > So, the autoencoder provides additional flexibility to our architectures, but comes with a challenge: the bracket structure we prescribe is enforced only in the latent space. This is an important point that could be clarified in the manuscript (e.g. limitation part) because most readers (including myself) expect this kind of model to respect physical laws in the original space in addition to the latent space through inductive biases.

Authorsrebuttal2023-08-16

We are glad that we could assuage your concerns. Thanks for being open to modifying the initial score. We will make sure to explicitly clarify the mentioned limitation in future versions of the paper.

Program Chairsdecision2023-09-21

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC