Towards Dynamic Message Passing on Graphs

Message passing plays a vital role in graph neural networks (GNNs) for effective feature learning. However, the over-reliance on input topology diminishes the efficacy of message passing and restricts the ability of GNNs. Despite efforts to mitigate the reliance, existing study encounters message-passing bottlenecks or high computational expense problems, which invokes the demands for flexible message passing with low complexity. In this paper, we propose a novel dynamic message-passing mechanism for GNNs. It projects graph nodes and learnable pseudo nodes into a common space with measurable spatial relations between them. With nodes moving in the space, their evolving relations facilitate flexible pathway construction for a dynamic message-passing process. Associating pseudo nodes to input graphs with their measured relations, graph nodes can communicate with each other intermediately through pseudo nodes under linear complexity. We further develop a GNN model named $\mathtt{\mathbf{N^2}}$ based on our dynamic message-passing mechanism. $\mathtt{\mathbf{N^2}}$ employs a single recurrent layer to recursively generate the displacements of nodes and construct optimal dynamic pathways. Evaluation on eighteen benchmarks demonstrates the superior performance of $\mathtt{\mathbf{N^2}}$ over popular GNNs. $\mathtt{\mathbf{N^2}}$ successfully scales to large-scale benchmarks and requires significantly fewer parameters for graph classification with the shared recurrent layer.

Paper

References (80)

Scroll for more · 38 remaining

Similar papers

Peer review

Reviewer UBFx6/10 · confidence 3/52024-07-06

Summary

This work proposes a new dynamic message passing method $N^2$, which initializes a number of pseudo nodes, and apply message passing across graph nodes and pseudo nodes. The message passing scheme relies on the proximity of the node embeddings, and the node embeddings are updated through message passing. The method is quite dynamic and flexible. Besides, it is empirically shown to alleviate oversquashing and oversmoothing problems. Experimental results show the significance of the proposed method.

Strengths

- Overall, the writing is good and clear, the concepts are precisely described, and the introduction of background is easily understandable. - The methodology is quite interesting, and the message passing design makes sense. - The experimental results are strong on most datasets. And the method is widely applicable to both graph and node level prediction. The ablations are carefully designed and the visualizations are pretty good. - The proposed method exhibits good performance on oversquashing and oversmoothing problems.

Weaknesses

- It would be good to have theoretical explanations why the method works well against oversquashing and oversmoothing problems, though not mandatory. - It would be good to compare with some graph structure learning or graph rewiring work. The baselines are basically GNNs and graph transformers.

Questions

- It is not intuitive to me why to use recurrent layers, except for the good property of weight sharing. Can you explain it? - Why use proximity measurement instead of some distance metrics?

Rating

6

Confidence

3

Soundness

3

Presentation

4

Contribution

3

Limitations

As mentioned by the authors, the performance deteriorates when the number of recurrent layers grow too large.

Reviewer 4Q2o5/10 · confidence 4/52024-07-07

Summary

The authors propose a dynamic message-passing scheme where graph nodes and learnable pseudo nodes are projected into a common space. This allows non adjacent nodes to communicate immediately whilst retaining linear complexity. The model performs well on a range of benchmarks and can reduce problems with over-squashing/over-smoothing compared to a baseline MPNN.

Strengths

- The novel (to my knowledge) method is explained in depth and well outlined. It is clearly explained and evaluated in the context of other approaches (rewiring, virtual node, transformers). - The results/experiments are really strong and the method is shown to help alleviate over-squashing/over-smoothing and performs well on a wide range of benchmarks. - Visualising the distribution of embedded nodes helps with understanding what is happening in practice and I really like this empirical understanding. - It is clear how this method lowers the computational complexity over methods such as Graph Transformers and the results indicate it can still outperform these approaches.

Weaknesses

- In line 45 the authors argue for their approach over using a single virtual node due to the fact that a VN can have a message-passing bottleneck. Firstly, I think this should be explained more rigorously rather than citing [1] as it is a key argument of the paper to not just have uniform global connections. Additionally, the extent to which your method alleviates this issue needs to be better explained. [Line 297], You use the balanced load argument to say that this helps alleviate the bottleneck. However, increasing connections between nodes within a community and not having global connections between communities would actually `increase' the bottleneck on those edges between the communities. The virtual node itself has less of a bottleneck BUT this does not mean bottlenecks in the graph are reduced. - On a similar note to above, we can reduce the bottleneck of the virtual node by subsampling edges (expander) or by adding more of them (this effectively increases their width) [1]. It is not clear in the paper how your dynamic and weighted approach improves over this (in terms of bottlenecks or some other property). - There are other dynamic message-passing schemes such as DRew [2]. The paper seems to argue for their approach over something like this due to pseudo nodes directly enabling global message-passing [line 82]. Given that these rewiring approaches seem relevant and closely-connected, I think this comparison and why your method can improve over this needs to be extended and evaluated in depth. For instance, in this case why would you need global message-passing in layer 1 when we always use > 1 layers? [1] Shirzad et al. EXPHORMER: Sparse Transformers for Graphs. ICML 2023. [2] Gutteridge et al. DRew: Dynamically Rewired Message Passing with Delay. ICML 2023.

Questions

- Do you use positional/structural encodings on these benchmarks? Does your baseline GIN/GCN + pseudo node also use these encodings? - It is not clear how the hyperparameter sweep is performed and what parameters are optimised over and on what metric. Could you provide some information on this? - Do you have any intuition why the method would outperform GTs on this benchmark? (not just have a lower computational complexity)

Rating

5

Confidence

4

Soundness

3

Presentation

3

Contribution

2

Limitations

some limitations are outlined in the appendix.

Reviewer ZLZY5/10 · confidence 3/52024-07-11

Summary

This paper proposes an adaptive message passing scheme for Graph Neural Networks that is based on learnable "pseudonodes" which, to a certain extent, decouple the paths along which node features are propagated from the topology of the underlying graph. Both pseudonodes and regular nodes in the underlying graph are embedded in a common space, which allows to utilize the common embeddings to generate proximity-dependent relations between nodes and pseudonodes that are used for a sparse "global" message passing layer. The proposed method is evaluated in a node and graph classification experiment for several benchmark data sets and the authors find superior performance on several data sets.

Strengths

[S1] The authors propose a new adaptive message passing scheme that introduces learnable pseudo-nodes and pseudo-edges, thus introducing dynamic pathways for message passing that are independent of the graph topology and that can be trained for a given learning task. The specific combination of pseudonde message passing and the use of recurrent layers is - to the best of my knowledge - new and original. [S2] The method is evaluated against several baseline methods for a node and graph classification task in 18 small and large-scale benchmark data sets. The experiments show superior performance for the proposed model in several of the data sets (all six data sets for graph classification, eight out of twelve data sets for node classification). [S3] Addressing limitations of GNNs that are due to over-squashing and over-smoothing, the authors address an important open issue in deep graph learning.

Weaknesses

[W1] I did not find the motivation to add additional learnable parameters to Graph Neural Networks that decouple message passing pathways from the topology of the input graph convincing. The authors motivate their work based on over-smoothing and over-squashing in GNNs but in my view the paper lacks an intuitive explanation why the proposed dynamic messing passing scheme should mitigate those problems, especially since the architecture additionally includes regular (local) message passing. To this end, I think that the different, complex local and global components of the dynamic message passing - though formally defined - are not explained well in the paper and some of the design choices appear to be rather arbitrary. [W2] I similarly could not follow the motivation for the addition of the recurrent layer, which the authors argue is added "to parameterize the displacements of embedded nodes" and to "revise the learned distribution [] of all embedded nodes and reshape the dynamic message passing pathways". A better explanation would be helpful. [W3] The idea to add pseudonodes for neural message passing has been previously explored, e.g. in the form of a sparse attention mechanism with so-called "global nodes" in the Exphormer architecture (https://arxiv.org/pdf/2303.06147). A better explanation of the contribution of the authors would be helpful, especially since the experiments show that the performance is very close to this architecture in many of the experiments. [W4] There are important details missing in the description of the experimental setup, namely whether hyperparameter tuning has been performed (i) for the proposed mod4el and (ii) for the baseline models. Moreover, in section B.2.2 the authors claim that "the detailed hyper-parameter settings on all benchmarks are reported in Tab. S6", however this table only includes hyperparameters for the proposed N2 model and not for baseline methods. I also checked the provided code, which does not cover the baseline experiments. As it is - despite the claims made in the answer to Q4 in the checklist - I do not consider the results showing superior performance compared to the baseline models reproducible based on the information provided in the paper and in the supplementary material. [W5] One could argue that the proposed approach to define pseudo-nodes and pseudo-edges that participate in the message passing could also be seen as a learnable graph pooling layer for GNNs, where pseudonodes take the role of supernodes. As such, I believe that the paper lacks a more detailed discussion of related works on (trainable) graph pooling operations (there is a single mention of graph pooling in the first paragraph of the related work section). [W6] Similarly, while the paper briefly mentions hierarchical Graph Neural Networks as related work that combines local and global message passing to learn multi-scale features, no explicit comparison to such approaches is included in the experimental evaluation.

Questions

I kindly ask the authors to address the following questions, which emerged during my review of the work: - Please provide more insights on the motivation behind the different ingredients of the dynamic message passing scheme, especially on (i) why the proposed local and global message scheme is supposed to address over-smoothing and over-squashing, and (ii) the motivation of the additional recurrent layer (see comments in [W1] and [W2]). For the first question, I see that there is an experimental analysis but I failed to see how this analysis supports the claims. For the latter question, it would be helpful to include an ablation study that removes the recurrent layer altogether. If I understood the ablation study correctly, currently only the influence of the number of recurrent layers is investigated. Or does the removal of the pseudo-code adaptation (Table 5) correspond to the removal of the recurrent layers? - Please clarify your contribution over other approaches using adaptive message passing architectures that combine local and non-local message exchanges (see comments in [W3], [W5] and [W6]). - Please provide details on the choice of hyperparameters and the use of hyperparameter tuning, both for the baseline methods and the proposed architecture (see detailed comments in [W3]) - Please add details on the ablation study results shown in Table 5, especially the number of experimental runs and the standard deviation of results. - Please increase the font of the results in tables 1 - 4. The text is too small to be readable in a printed version. Please also increase the size of the figures, especially figure 5 and 6, which are way too small to be readable in a printed version. During my review, I found the following typo: - line 259: Complex_i_ty Analysis

Rating

5

Confidence

3

Soundness

3

Presentation

2

Contribution

3

Limitations

In light of the open questions outlined above, I consider the discussion of limitations in appendix E overly short and not complete.

Area Chair b59X2024-08-07

anonymized scripts with experiment hyperparams

Dear reviewer, Take note that, as per the conference instructions, the authors forwarded me a link to an external anonymized page that shows the scripts they used & the associated hyperparameters. Let me know if you would like to see something specific and I will forward you the information. kind regards AC

Authorsrebuttal2024-08-13

Thank you very much for your positive feedback. We will refine our manuscript based on the rebuttal, including clarifying our motivations and contributions, delineating hyperparameter setups (based on our provided config files), and increasing the font size of tables and figures. We sincerely appreciate your suggestions, which have significantly helped us improve our work.

Reviewer Qj5n7/10 · confidence 5/52024-07-17

Summary

The paper considers the problem of flexible message passing with low complexity in GNNs. To tackle this concern, the paper proposes a novel dynamic message-passing mechanism for GNNs via projecting graph nodes and learnable pseudo nodes into a common space with measurable spatial relations. Based on this dynamic message passing mechanism, the paper constructs a GNN model Named N^2 to estimate the effectiveness and efficiency of the proposed message passing mechanism.

Strengths

1. A novel dynamic message passing mechanism for GNNs is proposed to provide flexible message passing with low complexity. The dynamic message passing is interesting. 2. The constructed N^2 GNN model is simple and effective. The complexity requirement is guaranteed. 3. The proposed N^2 holds superior performance.

Weaknesses

See questions.

Questions

1. The analysis of the dynamic message-passing mechanism. Can we learn an additional pseudo node to provide global guidance (like a cluster center)? 2. Can authors provide the discussion between dynamic and adaptive message passing? 3. Can authors show the possibility of incorporating the dynamic message passing mechanism into other networks?

Rating

7

Confidence

5

Soundness

4

Presentation

4

Contribution

4

Limitations

Yes.

Reviewer ZLZY2024-08-12

I thank the authors for the very detailed response. While I do not consider all of my criticism to be addressed fully, I do acknowledge that some of my questions on the motivation and the relation tyo graph pooling have been answered, which is why I decided to raise my score.

Reviewer 4Q2o2024-08-12

Thank you for your detailed response. The comparison to Drew and other rewiring based approaches is very compelling. I understand that your method can decrease pseudo-node bottlenecks but a better theoretical understanding of overcoming input-graph bottlenecks beyond the TreeMatch experiment is still lacking. The addition of a VN for the TreeMatch experiment does improve the paper and I have raised my initial score.

Authorsrebuttal2024-08-13

Thank you very much for your supportive feedback. We acknowledge that a theoretical understanding of how pseudo-node methods can detour messages from the input-graph bottlenecks is important. Due to time limitation, we provide two primary pathways to study this problem: - Measuring the curvature between graph nodes with and without pseudo nodes. - Given the messages to be passed between nodes i and j, recovering the messages from the pseudo-node pathway and the input-graph pathway to measure the information loss. For the first pathway, we performed an evaluation on amazon-ratings. By employing pseudo nodes, $N^2$ `creates message highways with positive curvature for 95% of graph-node pairs that are negatively curved on the input graphs.` Edges with `high negative curvature` cause the input-graph bottlenecks[R3]. This shows that **$N^2$ can overcome the input-graph bottlenecks by producing non-negatively curved message highways.** Your suggestions have enlightened us to further validate our methods from a theoretical perspective. We will keep studying this problem. [R3] Understanding over-squashing and bottlenecks on graphs via curvature. ICLR22

Reviewer Qj5n2024-08-14

Thank you for the detailed response. The authors make a clear explanation of my concerns. I retain my rating score.

Program Chairsdecision2024-09-25

Decision

Accept (poster)

© 2026 NYSGPT2525 LLC