Pith. sign in

REVIEW 5 major objections 6 minor 40 references

Fast and Distributed Equivariant Graph Neural Networks by Virtual Node Learning

T0 review · 5 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read This paper claims that a small ordered set of virtual nodes can replace most real-real message passing in equivariant graph neural networks, preserving accuracy while scaling to graphs with more than 100,000 nodes and cutting memory and…

desk verdict DistEGNN and the Fluid113K dataset are the real contributions here; the theoretical claim that virtual nodes approximate the real graph is shakier than the paper admits. read the letter →

arxiv 2506.19482 v1 pith:SQHOWSJL submitted 2025-06-24 cs.LG cs.AI

Yuelin Zhang , Jiacheng Cen , Jiaqi Han , Wenbing Huang This is my paper
classification cs.LGcs.AI
keywords equivariantgraphneuralnetworksvirtualnodesE(3)-equivariancesparsificationdistributedtrainingmaximummeandiscrepancyphysicaldynamicssimulationlarge-scalegeometricgraphs
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper is trying to establish that equivariant graph neural networks can be sparsified and distributed without losing accuracy by adding a small ordered set of virtual nodes. It argues that these virtual nodes, kept mutually distinct and spatially aligned with real nodes, can approximate the information carried by dense real-real message passing, so real-real edges can be dropped for speed. If true, it would let large physical simulation graphs, up to 113,000 nodes in the paper's Fluid113K benchmark, be processed on a single device or across multiple low-memory devices. The paper reports that FastEGNN beats full-edge EGNN across all three smaller benchmarks, and that DistEGNN cuts per-device memory sharply while limiting accuracy loss.

What carries the argument

The central object is the ordered set of $C$ virtual nodes with coordinates $\vec{Z}\in\mathbb{R}^{3\times C}$ and features $S\in\mathbb{R}^{H\times C}$, initialized at the graph center of mass so the model starts E(3)-equivariant. Three message types carry the computation: EGNN-style real-real messages $m^r_{ij}$ (Eq. 3), an E(3)-invariant virtual-global matrix $m^v = (\vec{Z} - \bar{x}\mathbf{1}^\top)^\top(\vec{Z} - \bar{x}\mathbf{1}^\top)$ (Eq. 4), and per-pair real-virtual messages $m^v_{ic}$ (Eq. 5). Real coordinates update by pulling toward neighbor differences and virtual-node differences, while virtual coordinates update by aggregating over all real nodes. The MMD loss in Eq. (10), with an RBF kernel, is the mechanism that keeps the virtual nodes spread out and aligned with the real distribution, and it is what makes Proposition IV.3's premise, that virtual coordinates approximate real ones, plausible.

What would settle it

Train the p=1.00 variant on the N-body system while sweeping C from 1 to N; if the MSE plateaus well above the full-edge EGNN baseline even when C=N and the virtual nodes are initialized at the real coordinates, the central approximation claim is falsified.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that a small ordered set of virtual nodes, each with a distinct semantic role, gives an E(3)-equivariant update $f(\vec{x}_i, \vec{Z}) = \vec{x}_i + \sum_{c=1}^C (\vec{z}_c - \vec{x}_i)\psi_c\big(\odot_{c=1}^C \|\vec{z}_c - \vec{x}_i\|^2, m^v\big)$ that is capable of approximating the real-node aggregation, provided the virtual coordinates $\vec{Z}$ approximate the real coordinates $\vec{X}$. The paper enforces that alignment with an E(3)-invariant Maximum Mean Discrepancy loss (Eq. 10), and enforces distinct roles by treating the virtual nodes as an ordered set with separate message-passing functions. FastEGNN realizes this as EGNN-style real-real message passing plus real-virtual message passing, and DistEGNN extends it by sharing the virtual nodes across partitioned devices and synchronizing them during training. The reported consequence is that with 75% of edges dropped, FastEGNN matches or beats full-edge EGNN, and with all edges dropped it degrades gracefully, whereas EGNN collapses to near-linear performance.

Load-bearing premise

The load-bearing premise is that a small, fixed set of virtual nodes whose coordinates match the real nodes' distribution can carry the information that full real-real message passing would have carried.

Editorial extensions

If this is right

  • Equivariant GNNs can be run on sparse graphs: dropping 75% of the longest edges costs little accuracy, with FastEGNN-⟨10,0.75⟩ reaching MSE 0.99×10⁻² on the N-body system, better than full-edge EGNN's 1.41×10⁻², at 53% of EGNN's inference time.
  • Even with all real-real edges removed, FastEGNN keeps physically meaningful predictions on Water-3D, with MSE 3.40×10⁻⁴ versus 12.38×10⁻⁴ for EGNN with all edges stripped, so virtual nodes alone can carry a substantial share of global information.
  • The virtual-node module transfers across architectures: FastRF, FastSchNet, and FastTFN improve over their baselines in accuracy and efficiency, with FastTFN-3-0.75 using 2% of TFN's memory and 20% of its time while lowering MSE by 56%.
  • DistEGNN makes 100K-node graphs tractable on modest hardware; with 8 devices, Fluid113K training is 7.34× faster and per-device memory drops from about 48 GB to 1.58 GB, while error only rises 15% across devices compared to 69% for EGNN.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • A direct extension of the paper's reasoning is that the virtual-node aggregator's capacity should scale with the complexity of the target interaction: for systems with strong many-body or long-range couplings, $C$ may need to grow with $N$, so a useful test is to measure the p=1.00 error as a function of $C/N$ on the N-body benchmark.
  • The same MMD-alignment mechanism could be reused for equivariant output heads beyond coordinates, such as forces or energies, where the virtual coordinates would serve as an invariant summary scaffold rather than as predicted positions; the paper does not test this.
  • Because DistEGNN's virtual nodes are synchronized across devices, communication cost is roughly linear in $C$ per layer regardless of partition count, suggesting the speedup over EGNN should keep growing as devices are added; this could be checked by profiling communication time separately from compute time.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 6 minor

Summary. The paper introduces FastEGNN and DistEGNN, two extensions of EGNN that add a small ordered set of learnable virtual nodes. FastEGNN keeps a sparse real-real edge set but connects every real node to every virtual node, using distinct message passing for each virtual channel and an MMD regularizer to encourage spatial alignment between virtual and real coordinates. DistEGNN partitions large graphs across devices and shares the virtual nodes across partitions to propagate global information. The authors report accuracy improvements over EGNN and other baselines on N-body, protein dynamics, Water-3D, and a new Fluid113K benchmark, along with substantial inference speedups when edges are dropped and per-device memory reductions in distributed training. The E(3)-equivariance proof in Appendix A is straightforward and appears correct. The central theoretical claim is that a small fixed set of virtual nodes can approximate the information carried by the full real-node graph, supported by Proposition IV.3, which is conditional on the virtual coordinates 'well approximating' the real coordinates. The main weaknesses are an inconsistency between the MMD formulation in Eq. (10) and the loss in Eq. (11), the lack of any estimate of variance across seeds, and an overstatement of the all-edges-dropped results on N-body.

Significance. If the claims hold, the paper advances the scalability of equivariant GNNs in a useful direction: it shows that a modest number of virtual nodes can preserve much of the expressivity of dense message passing while allowing aggressive edge sparsification, and it provides a concrete distributed scheme that processes graphs with over 100K nodes. The modular extension to RF, SchNet, and TFN broadens the potential impact. The new Fluid113K benchmark and the accompanied code release are concrete community assets. The E(3)-equivariance argument in Appendix A is clean and machine-verifiable in structure. However, the load-bearing premise that virtual nodes approximate the full real-node graph is not yet convincingly validated: the theoretical result is conditional, and the empirical evidence at the extreme edge-dropping rate is mixed, especially on N-body. The MMD target inconsistency and the absence of seed variance prevent the current empirical claims from being fully assessed.

major comments (5)
  1. [Section IV-C, Eq. (10)-(11)] There is a direct inconsistency between the MMD definition and its use in the loss. Eq. (10) defines LMMD using the predicted real node coordinates x_i^(L), but Eq. (11) writes LMMD(Z^(L), X^GT), i.e., against ground-truth final positions. If the implementation follows Eq. (11), then the regularizer leaks future target information during training while being absent at inference, so the alignment condition assumed by Proposition IV.3 is not guaranteed at test time. If the implementation follows Eq. (10), the paper should correct Eq. (11). The authors should state which target is used and provide an evaluation (e.g., measuring MMD between virtual coordinates and predicted real coordinates at inference) to support the approximation premise.
  2. [Section IV-D and Table I] The all-edges-dropped claims are overstated. The text says FastEGNN 'can still perform promisingly even when all edges are dropped' and that virtual nodes approximate the real graph, but Table I shows N-body MSE of 9.25–9.72e-2 at p=1.00, which is about 9x worse than the same model with full edges (1.04–1.10e-2) and only modestly better than the Linear baseline (12.66e-2). This contradicts the claim that a small ordered set of virtual nodes effectively approximates the large unordered graph of real nodes. The Water-3D and Protein results at p=1.00 are more supportive, but the N-body result should be discussed honestly and the claim should be restricted to settings where it actually holds.
  3. [Section VII-B, Tables I-III] No error bars, confidence intervals, or seed variance are reported anywhere in the empirical evaluation. Several accuracy differences used to support design choices are small (e.g., FastEGNN-⟨10,0.75⟩ at 0.99e-2 versus FastEGNN-⟨1,0.75⟩ at 1.03e-2 on N-body; the MMD ablation differences in Table II). Without multiple seeds or a variance estimate, the claimed superiority over baselines and the conclusions of the ablations cannot be reliably assessed. The authors should add seed variance or, at minimum, report the number of seeds and the standard deviation for the main tables.
  4. [Section IV-D, Proposition IV.3] Proposition IV.3 establishes a universal form for an update that depends only on virtual nodes, but it does not by itself prove that C virtual nodes can approximate the information from N real neighbors. The sentence 'f(x_i,Z) is able to universally approximate the messages from all the real nodes, if the virtual coordinates Z can well approximate the real ones X' is a heuristic bridge, not a theorem. The paper needs either a quantitative statement of approximation error as a function of C and the MMD, or a direct empirical test of the assumption (e.g., measuring the MMD between virtual and real node coordinates at inference on the test set). As written, the theoretical support for the central claim is not load-bearing.
  5. [Section VII-D, Tables IV-V] The distributed-memory claims should be qualified. While per-device memory is reduced, the total memory across devices grows with the number of devices, and DistEGNN often uses more total memory than EGNN at the same device count (e.g., Fluid113K at 8 devices: DistEGNN-5 total memory 12.62 GB versus EGNN 3.77 GB; relative training time 0.83 versus 0.15). The abstract's phrase 'dramatically reducing memory and computational overhead' is therefore only true per device and relative to the single-device setting. The paper should distinguish per-device and total memory, and should compare against the best feasible single-device baseline (which for Fluid113K is not available) rather than implying a universal overhead reduction.
minor comments (6)
  1. [Section IV-B, before Proposition IV.1] Typo: 'We have the flowing theoretical assurance' should read 'following'.
  2. [Section IV-B, Eq. (5)] The notation 'CM' in Eq. (5) is undefined; it appears to denote concatenation over c, but it should be written explicitly (e.g., concatenation symbol) and introduced in the text.
  3. [Appendix A, Theorem A.5] The phrase 'surjection from dao to' contains a typo; it should be 'surjection from ... to ...' with the intended sets spelled out.
  4. [Section VII-A and Table VIII] The dataset statistics for Water-3D are inconsistent: the text says 1000/100/100 trajectories, while Table VIII lists 15,000/1,500/1,500 samples. The authors should clarify the units and the relationship between trajectories and training samples.
  5. [General presentation] Several places contain duplicated or overlapping text blocks (e.g., around Eqs. (14)-(15) and in the captions of Figs. 3-4), likely artifacts of the LaTeX compilation. These should be cleaned before final submission.
  6. [Section VII-B1, observation 1] The claimed '29% improvement' over EGNN on N-body does not match Table I: 1.04e-2 versus 1.41e-2 is a 26% reduction. Please recompute or correct the percentage.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the core derivation is self-contained modulo standard cited lemmas, with one benign self-citation.

full rationale

The paper's central claim, that a small ordered set of virtual nodes can approximate the large real-node graph, is supported by Proposition IV.3, whose proof in the appendix relies on Lemma A.3 from Villar et al. [30] and Lemma A.4/Corollary 2 from the authors' earlier EGMN paper [29]. These are general equivariance theorems with stated assumptions that do not include the target virtual-node result, so citing them is legitimate independent support rather than circularity. The MMD loss in Eq. (11) is an auxiliary training regularizer, not a fitted prediction, and the paper's all-edges-dropped results honestly show that the virtual-node-only regime degrades sharply on N-body, so the empirical claim is not forced by construction. The only self-citation is the transparent footnote that this work extends the authors' ICML 2024 paper [11], and the distributed DistEGNN claims are validated against external EGNN baselines on independently constructed benchmarks. A genuine robustness gap exists: LMMD aligns virtual coordinates to ground-truth positions X^GT rather than to the model's own real coordinates X^(L), and Table IX shows it is estimated from only three sampled nodes on the smaller benchmarks, so the alignment premise of Proposition IV.3 is not guaranteed at inference. This is a correctness or generalization concern, not a circularity by construction, and therefore does not raise the circularity score.

Assumptions & free parameters 4 free parameters · 3 assumptions · 1 invented entities

The central claim rests on a few hyperparameters (C, p, sigma, lambda), on imported universal-approximation lemmas, and on the soft empirical assumption that virtual nodes can stand in for real-real message passing. No new physical entities are introduced; virtual nodes are internal model parameters trained by the objective.

free parameters (4)
  • Number of virtual nodes C = 1, 3, 10 (and 5 for Fluid113K)
    Swept per dataset; performance peaks at C=3 or C=10 depending on the dataset, showing sensitivity to this choice.
  • MMD bandwidth sigma (Eq. 10) = 1.5 (N-body), 1.0 (Protein), 1.5 (Water-3D), 3.0 (Fluid113K)
    Chosen per dataset in Table IX; controls the spatial scale of the RBF kernel in the MMD regularizer.
  • MMD balancing factor lambda (Eq. 11) = 0.03, 0.50, 0.01, 0.01
    Chosen per dataset; weight of the auxiliary MMD loss relative to the MSE loss.
  • Edge dropping rate p = 0.00, 0.75, 1.00
    Hyperparameter controlling sparsification; the claim of strong performance at p=1.00 only holds on Water-3D and Protein, not N-body.
assumptions (3)
  • standard math Equivariant universal-approximation lemmas from [29] and [30] hold as cited.
    Proposition IV.3 and Lemma A.4 depend on Corollary 2 of [29] and Proposition 10 of [30]; these results are imported without reproof.
  • domain assumption Virtual node coordinates, after MMD regularization, sufficiently approximate the real node distribution for Proposition IV.3 to apply.
    Section IV-D states that all-edges-dropped performance depends on the virtual coordinates being well aligned with the real coordinates; this alignment is only softly enforced by a regularizer and is empirically violated on N-body at p=1.00.
  • standard math Center-of-mass initialization of virtual nodes preserves E(3)-equivariance and permutation invariance of the updates.
    Eq. (2) and Proposition IV.1 rely on the CoM being equivariant to rotations and translations and invariant to permutations, a standard geometric fact.
invented entities (1)
  • Ordered learnable virtual nodes (coordinates Z and features S)
    purpose: Represent global context and substitute for real-real message passing on sparsified graphs; in DistEGNN they bridge subgraphs across devices.
    Virtual nodes are trained model parameters, not observable physical quantities. Their only handle is the MMD regularizer against real coordinates, which is internal to the paper's training objective.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fast and Distributed Equivariant Graph Neural Networks by Virtual Node Learning." pith.science (2026). https://pith.science/paper/SQHOWSJL

@misc{pith2026250619482,
  author       = {Pith},
  title        = {Pith review of: Fast and Distributed Equivariant Graph Neural Networks by Virtual Node Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SQHOWSJL}},
  note         = {Machine review of arXiv:2506.19482}
}
read the original abstract

Equivariant Graph Neural Networks (GNNs) have achieved remarkable success across diverse scientific applications. However, existing approaches face critical efficiency challenges when scaling to large geometric graphs and suffer significant performance degradation when the input graphs are sparsified for computational tractability. To address these limitations, we introduce FastEGNN and DistEGNN, two novel enhancements to equivariant GNNs for large-scale geometric graphs. FastEGNN employs a key innovation: a small ordered set of virtual nodes that effectively approximates the large unordered graph of real nodes. Specifically, we implement distinct message passing and aggregation mechanisms for different virtual nodes to ensure mutual distinctiveness, and minimize Maximum Mean Discrepancy (MMD) between virtual and real coordinates to achieve global distributedness. This design enables FastEGNN to maintain high accuracy while efficiently processing large-scale sparse graphs. For extremely large-scale geometric graphs, we present DistEGNN, a distributed extension where virtual nodes act as global bridges between subgraphs in different devices, maintaining consistency while dramatically reducing memory and computational overhead. We comprehensively evaluate our models across four challenging domains: N-body systems (100 nodes), protein dynamics (800 nodes), Water-3D (8,000 nodes), and our new Fluid113K benchmark (113,000 nodes). Results demonstrate superior efficiency and performance, establishing new capabilities in large-scale equivariant graph learning. Code is available at https://github.com/GLAD-RUC/DistEGNN.

Figures

Figures reproduced from arXiv: 2506.19482 by the authors.

Figure 1
Figure 1. The overall architecture of FastEGNN. (X⃗ , H) are the real coordinates and features; (Z⃗ , S) are the virtual coordinates and features. Each layer contains 5 components: Real-Real Message mr rirj , Virtual Global Message mv , Real-Virtual Message mv ri , Real Aggregation, and Virtual Aggregation. The real-real and real-virtual edges are displayed with different colors to indicate that the message passing and aggreg… view at source ↗
Figure 2
Figure 2. Overview of DistEGNN. Multiple devices collaboratively process a [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Rollout results of EGNN and FastEGNN-⟨10, 0.75⟩ (abbreviated as FE-⟨10, 0.75⟩ for brevity) on Water- EGNN exhibits unstable predictionswhereas our model produces results that match the ground truth more closely EGNN exhibits unstable predictions, whereas our model produces results that match the ground truth more closely. g. 3. oout esuts oGNN ad astGNN⟨0, 075⟩ (abbevated as ⟨0, 075⟩ obevty) oWate3 EGNN exhibits uns… view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: The results of different variants of FastEGNN on Protein Dynamics. We use green dots to indicate the learned locations of the virtual nodes. The Fig4The results of different variants of FastEGNN on Protein DynamicsWe use green dots to indicate the learned locations of …
Figure 5
Figure 5. Figure 5: The performance of four base and enhanced models on the Water-3D dataset in terms of speed, accuracy, and memory. Points closer to the lower-lef d lliiiditbttll ffii Fig. 5. The performance of four base and enhanced models on the Water-3D dataset in terms of speed, acc…
Figure 6
Figure 6. Figure 6: Visualization of inference time, accuracy, and memory usage of EGNN [ dlilithlt1234d 8 di Fig. 6. Visualization of inference time, accuracy, and memory usage of EGNN [ [PITH_FULL_IMAGE:figures/full_fig_p012_6.png]
Figure 7
Figure 7. Figure 7: Rollout results of EGNN and DistEGNN-5 on Fluid113K dataset, with training and inference performed in the distributed environment with 8 devices. Fig. 7. Rollout results of EGNN and DistEGNN-5 on Fluid113K dataset, with training and inference performed in the distribut…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 33 canonical work pages

  1. [1]

    Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,

    M. M. Bronstein, J. Bruna, T. Cohen, and P. Veli ˇckovi´c, “Geometric deep learning: Grids, groups, graphs, geodesics, and gauges,” arXiv preprint arXiv:2104.13478 , 2021

    arXiv 2021

  2. [2]

    A survey of geometric graph neural networks: data structures, models and applications,

    J. Han, J. Cen, L. Wu, Z. Li, X. Kong, R. Jiao, Z. Yu, T. Xu, F. Wu, Z. Wang et al. , “A survey of geometric graph neural networks: data structures, models and applications,” Frontiers of Computer Science , vol. 19, no. 11, p. 1911375, 2025

    work page 2025

  3. [3]

    A hitchhiker’s guide to geometric gnns for 3d atomic systems,

    A. Duval, S. V . Mathis, C. K. Joshi, V . Schmidt, S. Miret, F. D. Malliaros, T. Cohen, P. Lio, Y . Bengio, and M. Bronstein, “A hitchhiker’s guide to geometric gnns for 3d atomic systems,” arXiv preprint arXiv:2312.07511 , 2023

    arXiv 2023

  4. [4]

    Artificial intelligence for science in quantum, atomistic, and continuum systems,

    X. Zhang, L. Wang, J. Helwig, Y . Luo, C. Fu, Y . Xie, M. Liu, Y . Lin, Z. Xu, K. Yan, K. Adams, M. Weiler, X. Li, T. Fu, Y . Wang, A. Strasser, H. Yu, Y . Xie, X. Fu, S. Xu, Y . Liu, Y . Du, A. Saxton, H. Ling, H. Lawrence, H. Stärk, S. Gui, C. Edwards, N. Gao, A. Ladera, T. Wu, E. F. Hofgard, A. M. Tehrani, R. Wang, A. Daigavane, M. Bohde, J. Kurtin, Q....

    arXiv 2023

  5. [5]

    Equivariant spatio-temporal attentive graph networks to simulate physical dynamics,

    L. Wu, Z. Hou, J. Yuan, Y . Rong, and W. Huang, “Equivariant spatio-temporal attentive graph networks to simulate physical dynamics,” in Thirty-seventh Conference on Neural Information Processing Systems, 2023

    work page 2023

  6. [6]

    Equivariant graph neural operator for modeling 3d dynamics,

    M. Xu, J. Han, A. Lou, J. Kossaifi, A. Ramanathan, K. Az- izzadenesheli, J. Leskovec, S. Ermon, and A. Anandkumar, “Equivariant graph neural operator for modeling 3d dynamics,” in Forty-first International Conference on Machine Learning , 2024

    work page 2024

  7. [7]

    De novo design of protein structure and function with rfdiffusion,

    J. L. Watson, D. Juergens, N. R. Bennett, B. L. Trippe, J. Yim, H. E. Eisenach, W. Ahern, A. J. Borst, R. J. Ragotte, L. F. Milles et al., “De novo design of protein structure and function with rfdiffusion,” Nature, vol. 620, no. 7976, pp. 1089–1100, 2023

    work page 2023

  8. [8]

    Illuminating protein space with a programmable generative model,

    J. B. Ingraham, M. Baranov, Z. Costello, K. W. Barber, W. Wang, A. Ismail, V . Frappier, D. M. Lord, C. Ng-Thow- Hing, E. R. Van Vlack et al., “Illuminating protein space with a programmable generative model,” Nature, vol. 623, no. 7989, pp. 1070–1078, Nov 2023

    work page 2023

Show all 40 references
  1. [9]

    E (n) equiv- ariant graph neural networks,

    V . G. Satorras, E. Hoogeboom, and M. Welling, “E (n) equiv- ariant graph neural networks,” in International Conference on Machine Learning. PMLR, 2021, pp. 9323–9332

    2021

  2. [10]

    The fast multipole method: numerical implementa- tion,

    E. Darve, “The fast multipole method: numerical implementa- tion,” Journal of Computational Physics , vol. 160, no. 1, pp. 195–240, 2000

    2000

  3. [11]

    Improving equivariant graph neural networks on large geometric graphs via virtual nodes learning,

    Y . Zhang, J. Cen, J. Han, Z. Zhang, J. ZHOU, and W. Huang, “Improving equivariant graph neural networks on large geometric graphs via virtual nodes learning,” in Forty-first International Conference on Machine Learning , 2024

    2024

  4. [12]

    Integrating structured biological data by kernel maximum mean discrepancy,

    K. M. Borgwardt, A. Gretton, M. J. Rasch, H.-P. Kriegel, B. Schölkopf, and A. J. Smola, “Integrating structured biological data by kernel maximum mean discrepancy,” Bioinformatics, vol. 22, no. 14, pp. e49–e57, 2006

    2006

  5. [13]

    Schnet–a deep learning architecture for molecules and materials,

    K. T. Schütt, H. E. Sauceda, P.-J. Kindermans, A. Tkatchenko, and K.-R. Müller, “Schnet–a deep learning architecture for molecules and materials,” The Journal of Chemical Physics , vol. 148, no. 24, p. 241722, 2018

    2018

  6. [14]

    Directional message passing for molecular graphs,

    J. Klicpera, J. Groß, and S. Günnemann, “Directional message passing for molecular graphs,” in International Conference on Learning Representations, 2020

    2020

  7. [15]

    Tensor field networks: Rotation-and translation- equivariant neural networks for 3d point clouds,

    N. Thomas, T. Smidt, S. Kearnes, L. Yang, L. Li, K. Kohlhoff, and P. Riley, “Tensor field networks: Rotation-and translation- equivariant neural networks for 3d point clouds,” arXiv preprint arXiv:1802.08219, 2018

    2018 arXiv

  8. [16]

    Geometric and physical quantities improve e(3) equivariant message passing,

    J. Brandstetter, R. Hesselink, E. van der Pol, E. J. Bekkers, and M. Welling, “Geometric and physical quantities improve e(3) equivariant message passing,” in International Conference on Learning Representations, 2022

    2022

  9. [17]

    Se(3)- transformers: 3d roto-translation equivariant attention networks,

    F. Fuchs, D. Worrall, V . Fischer, and M. Welling, “Se(3)- transformers: 3d roto-translation equivariant attention networks,” in Annual Conference on Neural Information Processing Systems, vol. 33, 2020

    2020

  10. [18]

    Equivariant flows: sampling configurations for multi-body systems with symmetric energies,

    J. Köhler, L. Klein, and F. Noé, “Equivariant flows: sampling configurations for multi-body systems with symmetric energies,” arXiv preprint arXiv:1910.00753 , 2019

    1910 arXiv

  11. [19]

    Learning from protein structure with geometric vector perceptrons,

    B. Jing, S. Eismann, P. Suriana, R. J. L. Townshend, and R. Dror, “Learning from protein structure with geometric vector perceptrons,” in International Conference on Learning Representations, 2021

    2021

  12. [20]

    Are high- degree representations really unnecessary in equivariant graph neural networks?

    J. Cen, A. Li, N. Lin, Y . Ren, Z. Wang, and W. Huang, “Are high- degree representations really unnecessary in equivariant graph neural networks?” in Annual Conference on Neural Information Processing Systems, 2024

    2024

  13. [21]

    Neural message passing for quantum chemistry,

    J. Gilmer, S. S. Schoenholz, P. F. Riley, O. Vinyals, and G. E. Dahl, “Neural message passing for quantum chemistry,” in International Conference on Machine Learning , 2017

    2017

  14. [22]

    Graph classi- fication via deep learning with virtual nodes,

    T. Pham, T. Tran, H. Dam, and S. Venkatesh, “Graph classi- fication via deep learning with virtual nodes,” arXiv preprint arXiv:1708.04357, 2017

    2017 arXiv

  15. [23]

    An analysis of virtual nodes in graph neural networks for link prediction,

    E. Hwang, V . Thost, S. S. Dasgupta, and T. Ma, “An analysis of virtual nodes in graph neural networks for link prediction,” in The First Learning on Graphs Conference , 2022

    2022

  16. [24]

    Learning physical dynamics with subequivariant graph neural networks,

    J. Han, W. Huang, H. Ma, J. Li, J. B. Tenenbaum, and C. Gan, “Learning physical dynamics with subequivariant graph neural networks,” in Advances in Neural Information Processing Systems, 2022

    2022

  17. [25]

    Conditional antibody design as 3d equivariant graph translation,

    X. Kong, W. Huang, and Y . Liu, “Conditional antibody design as 3d equivariant graph translation,” in The Eleventh International Conference on Learning Representations , 2023

    2023

  18. [26]

    Graph neural networks: Scalability,

    H. Ma, Y . Rong, and J. Huang, “Graph neural networks: Scalability,” Graph Neural Networks: Foundations, Frontiers, and Applications, pp. 99–119, 2022

    2022

  19. [27]

    Scalable and effective graph neural networks via trainable random walk sampling,

    H. Ding, Z. Wei, and Y . Ye, “Scalable and effective graph neural networks via trainable random walk sampling,” IEEE Transactions on Knowledge and Data Engineering , vol. 37, no. 2, pp. 896–909, 2025

    2025

  20. [28]

    Distgnn: Scalable distributed training for large-scale graph neural networks,

    V . Md, S. Misra, G. Ma, R. Mohanty, E. Georganas, A. Heinecke, D. Kalamkar, N. K. Ahmed, and S. Avancha, “Distgnn: Scalable distributed training for large-scale graph neural networks,” in Proceedings of the International Conference for High Perfor- mance Computing, Networking...

    2021

  21. [29]

    Equivariant graph mechanics networks with constraints,

    W. Huang, J. Han, Y . Rong, T. Xu, F. Sun, and J. Huang, “Equivariant graph mechanics networks with constraints,” in International Conference on Learning Representations , 2022

    2022

  22. [30]

    Scalars are universal: Equivariant machine learning, structured like classical physics,

    S. Villar, D. W. Hogg, K. Storey-Fisher, W. Yao, and B. Blum- Smith, “Scalars are universal: Equivariant machine learning, structured like classical physics,” in Annual Conference on Neural Information Processing Systems , 2021

    2021

  23. [31]

    Metis—a software package for parti- tioning unstructured graphs, partitioning meshes and computing fill-reducing ordering of sparse matrices,

    G. Karypis and V . Kumar, “Metis—a software package for parti- tioning unstructured graphs, partitioning meshes and computing fill-reducing ordering of sparse matrices,” 01 1997

    1997

  24. [32]

    Neu- ral relational inference for interacting systems,

    T. Kipf, E. Fetaya, K.-C. Wang, M. Welling, and R. Zemel, “Neu- ral relational inference for interacting systems,” in International Conference on Machine Learning , 2018

    2018

  25. [33]

    Equivariant graph hierarchy-based neural networks,

    J. Han, W. Huang, T. Xu, and Y . Rong, “Equivariant graph hierarchy-based neural networks,” in Annual Conference on Neural Information Processing Systems , 2022

    2022

  26. [34]

    MDAnalysis: A Python Package for the Rapid Analysis of Molecular Dynamics Simulations,

    Richard J. Gowers, Max Linke, Jonathan Barnoud, Tyler J. E. Reddy, Manuel N. Melo, Sean L. Seyler, Jan Doma ´nski, David L. Dotson, Sébastien Buchoux, Ian M. Kenney, and Oliver Beckstein, “MDAnalysis: A Python Package for the Rapid Analysis of Molecular Dynamics Simulations,” ...

    2016

  27. [35]

    Molecular dynamics trajectory for benchmarking mdanalysis,

    S. Seyler and O. Beckstein, “Molecular dynamics trajectory for benchmarking mdanalysis,” p. m9, 2017

    2017

  28. [36]

    Learning to simulate complex physics with graph networks,

    A. Sanchez-Gonzalez, J. Godwin, T. Pfaff, R. Ying, J. Leskovec, and P. Battaglia, “Learning to simulate complex physics with graph networks,” in International Conference on Machine Learning. PMLR, 2020, pp. 8459–8468

    2020

  29. [37]

    La- grangian fluid simulation with continuous convolutions,

    B. Ummenhofer, L. Prantl, N. Thuerey, and V . Koltun, “La- grangian fluid simulation with continuous convolutions,” in International Conference on Learning Representations , 2020

    2020

  30. [38]

    SPlisHSPlasH Library

    J. Bender et al., “SPlisHSPlasH Library.” [Online]. Available: https://github.com/InteractiveComputerGraphics/SPlisHSPlasH 17 APPENDIX A. Proof Theorem A.1 (Propostion IV .1). If the initialization of the virtual nodes satisfies Eq. (2), then after Eqs. (3) to (9), the output ...

  31. [39]

    AMONG THEM , F LUID 113K IS OUR GENERATED LARGE -SCALE FLUID SIMULATION DATASET , WHERE EACH GRAPH CONTAINS OVER 100K NODES AND AN AVERAGE OF 1.7M EDGES

    Dataset Details TABLE VIII BASIC STATISTICS OF THE FOUR DATASETS , WITH GRAPH SIZES INCREASING PROGRESSIVELY . AMONG THEM , F LUID 113K IS OUR GENERATED LARGE -SCALE FLUID SIMULATION DATASET , WHERE EACH GRAPH CONTAINS OVER 100K NODES AND AN AVERAGE OF 1.7M EDGES . N-body Syst...

  32. [40]

    (10) 1.5 1 .0 1 .5 3 .0 Balancing factor λ in Eq

    Implementation Details TABLE IX DEFAULT HYPER -PARAMETERS ON FOUR DATASETS Hyperparameter N-body System Protein Dynamics Water-3D Fluid-113K Bandwidth σ in Eq. (10) 1.5 1 .0 1 .5 3 .0 Balancing factor λ in Eq. (11) 0.03 0 .50 0 .01 0 .01 Number of Sampled Nodes N in Eq. (10) 3...

    2000

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.