Benchmarking Sheaf Neural Networks for Inductive Tasks

Edoardo Coppola; Pietro Li\`o; Stefano Fiorini

REVIEW 4 major objections 6 minor 22 references

Across 1,890 controlled inductive experiments, the parameterization of edge restriction maps—not the diffusion mechanism—dominates sheaf neural network performance, with general unconstrained maps beating diagonal and orthogonal ones at eve

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 04:47 UTC pith:5GXGRH7C

load-bearing objection First inductive SNN benchmark with a clear, useful result—but the Laplacian-free reformulation's fidelity to published NSD/SAN is asserted, not checked, and code is withheld; worth a serious referee with conditions. the 4 major comments →

arxiv 2608.02558 v1 pith:5GXGRH7C submitted 2026-08-03 cs.LG

Benchmarking Sheaf Neural Networks for Inductive Tasks

Stefano Fiorini , Edoardo Coppola , Pietro Li\`o This is my paper

classification cs.LG

keywords sheaf neural networksinductive graph learningrestriction mapsgraph neural networksmessage passinggraph benchmarksstalk dimensiongeneralization

verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper asks what actually drives performance in Sheaf Neural Networks when they are trained and tested on unseen graphs. It runs a large factorial benchmark, 1,890 controlled runs across 14 inductive datasets, and finds that the choice of restriction-map parameterization (the per-edge linear maps that transport information between node stalks) dominates, explaining about 12% of score variance, while the choice of diffusion mechanism explains only 0.2%. General unconstrained maps consistently outperform diagonal and orthogonal maps at every stalk dimension. Larger stalk dimensions help, but behave as generic added capacity—improvements are uncorrelated with graph size or degree and are just as large on long-range benchmarks—rather than a targeted fix for long-range information flow. The paper concludes that a single sheaf configuration generalizes across datasets, that tuning the surrounding architectural recipe is where remaining gains lie, and that SNNs do transfer to inductive settings but remain below the strongest baselines, with dataset-dependent gaps.

Core claim

The principal finding is that, under a matched inductive protocol, the sheaf-specific design axis that matters is the restriction-map family: Gen > Diag > O(d) holds in all six mechanism-by-stalk-dimension slices, with the Gen–O(d) gap roughly constant (~0.23 in rank-normalized score) as stalk dimension grows, indicating that the orthogonal constraint (norm preservation) rather than parameter count limits orthogonal maps. The diffusion mechanism—sheaf diffusion, static sheaf attention, or the new dynamic-attention variant—has negligible aggregate effect at 0.2% variance, though the two main mechanisms trade rankings depending on whether the full component recipe is used. The surrounding GNN

What carries the argument

The enabler is a Laplacian-free, edge-wise message-passing reformulation of sheaf diffusion: instead of assembling the n·d × n·d sheaf Laplacian, each edge computes a message that routes the neighbor's representation through the per-edge transport block P_ij = F_i^T F_j (the product of the two endpoint restriction maps), with an optional attention gate, and a node-wise update. This keeps time and memory linear in nodes and edges, and composes with standard cross-graph batching, making the full sheaf design space trainable under inductive protocols for the first time. The three restriction-map families (diagonal, orthogonal O(d), and general unconstrained matrices) are the axes along which th

Load-bearing premise

The whole benchmark rests on the claim that the new edge-wise message-passing operator computes the same thing as the original sheaf-Laplacian diffusion; the paper does not verify this equivalence numerically or formally, so if the reformulation changes normalization, gating, or gradient flow, the results describe the new variant rather than the published sheaf models.

What would settle it

Compare the reformulated operator against the explicit construction of the normalized sheaf Laplacian on a small graph with fixed restriction maps: if outputs or training gradients differ beyond numerical tolerance, the equivalence claim on which the benchmark is built fails. Alternatively, rerun the full factorial with the original Laplacian-based implementation on a single fixed graph and check whether the ordering Gen > Diag > O(d) and the 12.2% vs 0.2% variance attribution persist.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

Sheaf diffusion can be trained and evaluated on batches of unseen graphs without ever materializing the heavy sheaf Laplacian, so the full sheaf design space is now accessible to inductive and large-scale protocols.
Inductive sheaf network design should prioritize the restriction-map family over the diffusion mechanism; general (unconstrained) maps are the safe default.
Increasing stalk dimension is a capacity knob, not a long-range-reach knob; it does not reduce the need for architectural mechanisms that actually distribute information across distant nodes.
A single sheaf configuration (e.g., general maps at stalk dimension 4) performs near-optimally across diverse datasets, so per-dataset tuning budgets are better spent on the component recipe (residuals, edge features, normalization) than on the sheaf operator.
On matched budgets and protocol, current sheaf models transfer to inductive settings but do not reach the strongest published baselines; the gap is dataset-dependent and can be as small as about 1.4% on some node-level benchmarks and much larger on others.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

If the edge-wise reformulation is not exactly equivalent to the published sheaf-Laplacian diffusion—no formal or numerical equivalence check is reported—the benchmark's conclusions characterize the reformulated operator rather than the original models; a numerical equivalence check on a small graph would settle this.
The BatchNorm failure on ogbg-ppa, where validation collapses while training fits, points to a train/test statistic mismatch that may also affect other architectures under the same fixed protocol; testing alternative normalization schemes (e.g., layer norm) is a natural next step.
The near-parity on datasets where baselines themselves differ little suggests the sheaf gap may be bounded by the protocol's headroom; comparing sheaf models against ablations with weaker baselines would clarify whether the gap is intrinsic to sheaf diffusion.
The finding that the component recipe dominates the sheaf-specific design predicts that simply adding sheaf restriction maps to an already strong recipe yields little beyond the recipe itself—a testable hypothesis for future architecture search.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

First inductive SNN benchmark with a clear, useful result—but the Laplacian-free reformulation's fidelity to published NSD/SAN is asserted, not checked, and code is withheld; worth a serious referee with conditions.

read the letter

Two things to know about arXiv:2608.02558. First, it delivers the first systematic inductive evaluation of sheaf neural networks: 1,890 controlled runs across 14 datasets, a two-stage screen-then-confirm protocol, rank-normalized scores with exact Wilcoxon tests, and a disclosed anomaly on ogbg-ppa. The headline result is plausible and actionable: restriction-map parameterization drives performance (12.2% variance) while diffusion mechanism is nearly inert (0.2%), and the surrounding GNN+ recipe explains more variance than the entire sheaf design space. Second, the paper's central implementational claim—that the Laplacian-free edge-wise reformulation reproduces the published NSD/SAN layers—is not verified by any numerical or formal check, and code is withheld until acceptance. That is the soft spot that matters.

What's good: the authors inherit the GNN+ pipeline verbatim, use matched parameter budgets, report test-at-best-validation, include negative and anomalous results, and use paired statistics rather than per-cell cherry-picking. The ogbg-ppa BatchNorm collapse is reported with an explanation rather than hidden. SANv2 is a modest but clearly labeled extension (GATv2-style dynamic scoring) and the paper does not oversell it.

Soft spots, in proportion: (1) The reformulation equivalence. Section 4 gives a derivation sketch; for NSD it is plausible—the math works out for the symmetric normalized Laplacian action, though the W1/W2 mixing is not shown—but for SAN/SANv2 the channel-wise attention gate (Γ=diag(α)) is not the scalar lifted attention of the original SAN (Λ̂=Λ⊗1_d). That is a deliberate variant, yet the paper calls it "SAN" and builds the NSD-vs-SAN comparison on it. If the attention weighting differs, that comparison may not transfer to the published SAN. The authors should at least provide a numerical equivalence check on small graphs. (2) Single-seed screening. The design-space conclusions rest on one seed per cell. The paired Wilcoxon tests over 14 datasets give some protection, and the restriction-map ordering is stable across all six mechanism×d slices, so the direction is likely right; the η² magnitudes, however, should be read as approximate. (3) Baselines are quoted, not re-run. Given the inherited pipeline and matched budget, this is acceptable, but it does put weight on environment parity.

Who this is for: anyone working on sheaf GNNs or on fair GNN benchmarking. The practical guidance—spend tuning budget on the recipe, not the sheaf operator—is useful even if the fidelity concern holds, because the paper shows that within this large design space the recipe dominates.

Recommendation: send it to peer review. The benchmark is important and the reporting is honest. Referees should require code release or a reproducibility appendix, a numerical equivalence check for the Laplacian-free reformulation, and ideally a few multi-seed replicates for key cells. These are addressable, not fatal.

Referee Report

4 major / 6 minor

Summary. The paper presents a large-scale inductive benchmark of sheaf neural networks (SNNs). It reformulates NSD and SAN as Laplacian-free, edge-wise message-passing operators, introduces a new SANv2 mechanism with GATv2-style dynamic attention, and evaluates a factorial design mixing three diffusion mechanisms, three restriction-map parameterizations, three stalk dimensions, and six GNN+ architectural components across 14 inductive datasets. The screening phase comprises 1,890 single-seed runs analyzed with rank-normalized scores and exact Wilcoxon tests; selected configurations are confirmed at full budget with five seeds. The authors conclude that restriction-map parameterization dominates the diffusion mechanism, larger stalk dimension adds capacity rather than long-range reach, the surrounding GNN+ recipe explains more variance than the entire sheaf design space, and SNNs transfer to inductive settings but do not match the strongest baselines.

Significance. If its central claims hold, this is the first systematic inductive characterization of SNNs and would substantively redirect research effort from diffusion mechanisms toward restriction maps and architectural recipes. The paper has genuine strengths: a large and carefully controlled factorial design, exact paired statistics on rank-normalized scores, explicit disclosure of anomalies such as the ogbg-ppa BatchNorm collapse, a screen-then-confirm protocol, and a matched GNN+ pipeline inherited from Luo et al. (2025). However, the benchmark's external validity rests on the faithfulness of the §4 Laplacian-free reformulation to the published NSD/SAN layers, and the main design-space conclusions are computed from single-seed screening runs with no seed-variance analysis. Both points need to be addressed before the headline conclusions can be taken as established.

major comments (4)

[§4, Eqs. (4)–(8)] The Laplacian-free operator is asserted to be a reformulation of published NSD and SAN, but the equations do not match the reference layers. NSD in Eq. (2) applies Δ_F to (I_n⊗W1)XW2; the implementation in Eq. (7) normalizes by D_i^{-1/2} and does not show W1/W2 mixing. SAN in Eq. (3) uses scalar lifted attention Λ⊗1_d; Eq. (8) instead uses a channel-wise multi-head gate diag(α_ij) and a high-pass self-term. No numerical or formal equivalence check is provided, and §8.1 states code will be released only upon acceptance. Since every design-space conclusion in §6.1 is computed over the mechanisms implemented by these equations, the benchmark may characterize new variants rather than NSD/SAN. Please provide an explicit derivation or a small equivalence test on random graphs, and make the code available for review.
[§6.1, Table 1; §8.4] All screening conclusions — including the headline ordering Gen>Diag>O(d), the stalk-dimension monotonicity, and the η² values — are based on a single seed per configuration. Exact Wilcoxon tests are appropriate for paired differences, but the paired differences are single draws; no seed variance is reported. The statement in §8.4 that screening 'uses a single-design' because it targets aggregate trends does not address the risk that one outlier seed can change ranks and hence the variance decomposition. The authors should run at least a stratified subset of cells with 2–3 seeds and show that the aggregate rankings and η² values are stable, or otherwise bound the seed-noise contribution.
[§6.1, Table 1] The η² values (restriction map 12.2%, diffusion mechanism 0.2%, GNN+ component 26.7%, full sheaf space 21.1%) are central to the paper's message, but the variance-decomposition methodology is not specified. No equation is given for how η² is computed on within-dataset rank-normalized scores with dataset as the unit of analysis, nor whether interactions are included. Without a precise definition (e.g., SS_effect/S_total from a specified model), these headline percentages are not auditable. Please state the exact decomposition and the underlying model.
[§6.1, ogbg-ppa] ogbg-ppa is retained in all 14-dataset paired comparisons and variance decompositions even though the authors report that all sheaf models collapse to near chance under the prescribed BatchNorm. The paper says this 'does not distort aggregate conclusions,' but no leave-one-out sensitivity analysis is shown. One near-constant dataset can strongly affect rank normalization and variance shares. Please report the key comparisons (Gen vs O(d), d=4 vs d=2, diffusion mechanism, architecture-vs-sheaf η²) with ogbg-ppa excluded.

minor comments (6)

[§5, Datasets and Tasks] The paragraph 'Datasets and Tasks' is duplicated: 'We evaluate sheaf-based models on 14 inductive datasets...' appears twice, as does the sentence about official splits. Please remove the repetition.
[§4, Eqs. (7)–(8)] The notation for the mixed representation is inconsistent: Eq. (7) does not define W1/W2 in the NSD path, while Eq. (8) introduces ilde{x} without explicitly stating its shape. Please define ilde{x}=W1 x W2^⊤ and state whether the NSD path includes the same mixing.
[Table 1] The caption says 'averaged over 12 datasets,' but the table rows list 13/14 datasets favoring a direction and the text refers to 14 datasets. Please reconcile the number of datasets used in the mean Δ computation.
[Table 2] On CLUSTER, SANv2 reports 0.6805±0.0678, a standard deviation more than an order of magnitude larger than the other entries. This suggests seed instability or a failed run; it should be commented on or investigated.
[§6.1] The phrases 'only0/14and2/14datasets improving' and '0.316onthefoursuperpixeldatasetsagainst0.395onOGB' suffer from missing spacing/formatting; please fix.
[§5, Baselines] The paper compares against quoted baseline numbers rather than re-running them. This is acceptable given the inherited pipeline, but the text should state more explicitly that small hardware/software differences could affect comparability beyond the matched protocol.

Circularity Check

0 steps flagged

Benchmark study with no circular derivation; central claims are empirical comparisons against external baselines. Minor self-citations exist but are not load-bearing.

full rationale

This is a benchmarking paper, not a derivation of predictions from fitted parameters. The central claims (restriction-map parameterization dominates, diffusion mechanism has negligible effect, stalk size adds capacity rather than range) are descriptive statistics over a 1,890-run factorial against an inherited, externally published training pipeline and external baseline numbers. No equation in the paper reduces to an input by construction: the eta-squared values are summaries of measured scores, not quantities fitted to the conclusions they support. The paper explicitly acknowledges the one 'by construction' element — per-dataset selection attaining rank-normalized score 1.0 — and then checks the more interesting global configuration, which is a legitimate robustness analysis rather than circularity. 'SANv2 inherits the best SAN configuration' is a controlled comparison design, not a fitted-input-called-prediction step. The main verification gap is that the Laplacian-free reformulation (Eqs. 4–8) is asserted to reproduce published NSD/SAN layers without a numerical or formal equivalence check, and code is withheld until acceptance. That is a correctness/external-validity concern, not circularity: the paper's conclusions do not define the reformulation in terms of those conclusions, and the published definitions from Bodnar et al. (2022) and Barbero et al. (2022b) are used as background, not as a load-bearing self-citation chain. No uniqueness theorem from the authors is invoked, and no renamed empirical pattern is presented as a derivation. Overall, no significant circularity; score is low because of mild self-citation in background and the unverified-but-asserted reformulation, neither of which drives the main findings circularly.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

This is an empirical benchmark rather than a derivation, so the ledger mainly records protocol assumptions and experimental design choices. The central claims rest on equivalence of the batching reformulation to reference SNNs, on the inherited pipeline making quoted baselines fair, and on the aggregate statistical treatment of 14 datasets. No free parameters are fitted to make a prediction; the design axes are the object of study.

free parameters (4)

Stalk dimension d = d ∈ {2,3,4}
Explored as a design axis. The claim that larger stalks add capacity rather than long-range reach is an empirical trend over these three chosen values, not a law.
Channels-per-stalk f and layer width = Set per dataset so c = d·f fits the Luo et al. parameter budget (Table 10)
Confounds stalk dimension with total per-node capacity: d=4 models use fewer channels f, so the 'capacity' interpretation depends on how this trade-off is resolved.
Layer count L and hidden width per dataset = Per-dataset values in Table 10
Confirmation performance depends on these validation-selected hyperparameters; the 'single global sheaf config' claim fixes the sheaf block but tunes the recipe per dataset.
Attention heads H = 8
Fixed by hand for SAN/SANv2; no sensitivity analysis is reported.

axioms (5)

domain assumption The message-passing reformulation (Eqs. 4–8) computes the same update as the sheaf Laplacian diffusion (Eqs. 2–3).
§4 introduces the reformulation; equivalence to reference NSD/SAN is asserted, not verified, and is the basis for attributing benchmark results to SNNs.
domain assumption The inherited GNN+ training pipeline and matched parameter budget from Luo et al. (2025) make quoted baseline numbers directly comparable.
§8.3 and §8.6; baselines are not re-run, so the 'do not reach strongest baselines' claim rests on this inherited-protocol assumption.
domain assumption Rank-normalized within-dataset scores and paired Wilcoxon tests across 14 datasets are a valid way to aggregate claims across benchmarks.
§6.1; with n=14 the power of exact Wilcoxon tests is limited, and the variance-decomposition percentages (η²) are not fully specified.
standard math Standard numerical linear algebra (Householder reflections for O(d), Newton-Schulz for D^{-1/2}, tanh-bounded map learner) behaves as specified.
§10; these are standard methods with no novel mathematical proofs, but their numerical behavior is load-bearing for the experiments.
ad hoc to paper Including ogbg-ppa, where sheaf models collapse to near chance because of BatchNorm, does not distort aggregate design-space conclusions.
§6.1 describes the anomaly but retains the dataset in all paired tests; the paper treats this as conservative, but it is a protocol choice that affects the aggregate statistics.

invented entities (1)

SANv2 (sheaf attention with GATv2-style dynamic scoring) no independent evidence
purpose: Adds query-dependent attention scoring to sheaf attention, evaluated against static-attention SAN.
Introduced and evaluated only within this paper; it does not consistently outperform SAN, and there is no external benchmark or formal characterization.

pith-pipeline@v1.3.0-daily-deepseek · 27761 in / 15928 out tokens · 171573 ms · 2026-08-04T04:47:27.865815+00:00 · methodology

0 comments

read the original abstract

Sheaf Neural Networks (SNNs) generalize message passing by replacing scalar edge weights of standard Graph Neural Networks (GNNs) with learnable, edge-dependent restriction maps between node stalks. Despite their strong theoretical foundations and promising transductive results, SNNs have been evaluated almost exclusively on transductive node classification, leaving their behaviour under inductive protocols unknown. We address this gap through the first systematic benchmark of the sheaf design space, evaluating three diffusion mechanisms (neural sheaf diffusion, sheaf attention, and sheaf attention with Graph Attention Network v2), three restriction-map parameterizations, three stalk dimensions, and six modern GNN architectural components, within a message-passing reformulation that never assembles the heavy sheaf Laplacian, making the full design space trainable under cross-graph batching. Across $1{,}890$ controlled experiments on 14 inductive datasets, multiple insights emerge: restriction maps are the dominant design choice and general maps are preferable, larger stalks add capacity but not long-range reach, architectural components explain more performance variation than the entire sheaf-specific design space itself. Under a matched protocol, SNNs transfer to inductive settings but do not reach the strongest baselines, with gaps being dataset-dependent. Practically, a single sheaf configuration can generalize across datasets, so effort is better spent tuning the surrounding architectural recipe than the sheaf operator itself.

Figures

Figures reproduced from arXiv: 2608.02558 by Edoardo Coppola, Pietro Li\`o, Stefano Fiorini.

**Figure 1.** Figure 1: Benchmarking workflow overview. (Top-left) sheaf design space spanning 3 diffusion mechanisms, 3 restrictionmap parameterizations, and 6 GNN components; (top-right) our Laplacian-free inductive message passing; (bottom) evaluation benchmarks, datasets, and domain taxonomy. 2020). This richer diffusion mechanism has been shown, both theoretically and empirically, to improve robustness to heterophily and mi… view at source ↗

**Figure 2.** Figure 2: Screening marginal effects. Each point is the mean [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** Figure 3: Complete screening grid, GNN Benchmark. Rows are GNN [PITH_FULL_IMAGE:figures/full_fig_p014_3.png] view at source ↗

**Figure 4.** Figure 4: Complete screening grid, LRGB. Rows are GNN [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗

**Figure 5.** Figure 5: Complete screening grid, OGB. Rows are GNN [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗

discussion (0)

Reference graph

Works this paper leans on

22 extracted references · 16 linked inside Pith

[4]

Brody, S.; Alon, U.; and Yahav, E

Heterogeneous sheaf neural networks.arXiv preprint arXiv:2409.08036. Brody, S.; Alon, U.; and Yahav, E

Pith/arXiv arXiv
[5]

We do not use mixed precision, and all reported numbers are produced in float32. Component Version Python 3.10.20 PyTorch 2.2.0 (+cu118) CUDA toolkit 11.8 PyTorch Geometric 2.3.1 pyg-lib0.4.0 torch-scatter2.1.2 torch-sparse0.6.18 torch-cluster1.6.3 torchmetrics1.9.0 GPU NVIDIA A100 40–80GB Table 5: Software and hardware environment. The com- plete depende...

2025
[6]

Di Giovanni, F.; Rowbottom, J.; Chamberlain, B

Topology-Informed Graph Transformer.arXiv preprint arXiv:2402.02005. Di Giovanni, F.; Rowbottom, J.; Chamberlain, B. P.; Markovich, T.; and Bronstein, M. M

arXiv
[8]

Errica, F.; Podda, M.; Bacciu, D.; and Micheli, A

Long range graph benchmark.arXiv preprint arXiv:2206.08164. Errica, F.; Podda, M.; Bacciu, D.; and Micheli, A

Pith/arXiv arXiv
[11]

arXiv preprint arXiv:2012.06333

Sheaf neural networks. arXiv preprint arXiv:2012.06333. Hansen, J.; and Ghrist, R

Pith/arXiv arXiv 2012
[12]

InInternational conference on machine learning, 12724–12745

A generalization of vit/mlp-mixer to graphs. InInternational conference on machine learning, 12724–12745. PMLR. Hu, W.; Fey, M.; Zitnik, M.; Dong, Y.; Ren, H.; Liu, B.; Catasta,M.;andLeskovec,J.2020. Opengraphbenchmark: Datasetsformachinelearningongraphs.Advancesinneural information processing systems, 33: 22118–22133. Huang, Y.; Miao, S.; and Li, P

2020
[13]

Hussain, M

What Can We Learn from State Space Models for Machine Learning on Graphs? arXiv preprint arXiv:2406.05815. Hussain, M. S.; Zaki, M. J.; and Subramanian, D

Pith/arXiv arXiv
[14]

InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 655–665

Globalself-attentionasareplacementforgraphconvolution. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 655–665. Kipf,T.N.;andWelling,M.2017.Semi-SupervisedClassifi- cationwithGraphConvolutionalNetworks. InInternational Conference on Learning Representations. Li, Q.; Han, Z.; and Wu, X.-M

2017
[15]

Luo, Y.; Shi, L.; and Wu, X.-M

Graph Ex- ternal Attention Enhanced Transformer.arXiv preprint arXiv:2405.21061. Luo, Y.; Shi, L.; and Wu, X.-M

Pith/arXiv arXiv
[16]

Ma, L.; Lin, C.; Lim, D.; Romero-Soriano, A.; Dokania, P

Can classic gnns be strong baselines for graph-level tasks? simple architectures meet excellence.arXiv preprint arXiv:2502.09263. Ma, L.; Lin, C.; Lim, D.; Romero-Soriano, A.; Dokania, P. K.; Coates, M.; Torr, P.; and Lim, S.-N

arXiv
[17]

arXiv preprint arXiv:2305.17589

Graph Inductive Biases in Transformers without Message Passing. arXiv preprint arXiv:2305.17589. Rampášek, L.; Galkin, M.; Dwivedi, V. P.; Luu, A. T.; Wolf, G.; and Beaini, D

Pith/arXiv arXiv
[18]

Ribeiro, A.; Tenório, A

Recipe for a General, Powerful, Scalable Graph Transformer.arXiv preprint arXiv:2205.12454. Ribeiro, A.; Tenório, A. L.; Belieni, J.; Souza, A. H.; and Mesquita, D

Pith/arXiv arXiv
[19]

arXiv preprint arXiv:2507.00647

Cooperative Sheaf Neural Networks. arXiv preprint arXiv:2507.00647. Sancak, K.; Hua, Z.; Fang, J.; Xie, Y.; Malevich, A.; Long, B.;Balin,M.F.;andÇatalyürek,Ü.V.2024. AScalableand Effective Alternative to Graph Transformers.arXiv preprint arXiv:2406.12059. Shirzad, H.; Velingker, A.; Venkatachalam, B.; Sutherland, D.J.;andSinop,A.K.2023. Exphormer:Sparsetr...

arXiv 2024
[20]

Traversa,P.;FerrazdeArruda,G.;Vazquez,A.;andMoreno, Y.2023.Robustnessandcomplexityofdirectedandweighted metabolic hypergraphs

Where did the gap go? reassessing the long-range graph benchmark.arXiv preprint arXiv:2309.00367. Traversa,P.;FerrazdeArruda,G.;Vazquez,A.;andMoreno, Y.2023.Robustnessandcomplexityofdirectedandweighted metabolic hypergraphs. Veličković,P.;Cucurull,G.;Casanova,A.;Romero,A.;Liò, P.; and Bengio, Y

Pith/arXiv arXiv 2023
[21]

Wu,S.;Sun,F.;Zhang,W.;Xie,X.;andCui,B.2022

Graph- mamba: Towards long-range graph sequence modeling with selective state spaces.arXiv preprint arXiv:2402.00789. Wu,S.;Sun,F.;Zhang,W.;Xie,X.;andCui,B.2022. Graph neural networks in recommender systems: a survey.ACM computing surveys, 55(5): 1–37. 8 Experimental Setup and Reproducibility 8.1 Code The code associated with this work will be released up...

Pith/arXiv arXiv 2022
[2019]

Fiorini, S.; Aktas, H.; Duta, I.; Coniglio, S.; Morerio, P.; Del Bue, A.; and Liò, P

A fair comparison of graph neural networks for graph classifi- cation.arXiv preprint arXiv:1912.09893. Fiorini, S.; Aktas, H.; Duta, I.; Coniglio, S.; Morerio, P.; Del Bue, A.; and Liò, P

Pith/arXiv arXiv 1912
[2020]

Bar-Shalom, G.; Bevilacqua, B.; and Maron, H

On the bottleneck of graph neuralnetworksanditspracticalimplications.arXivpreprint arXiv:2006.05205. Bar-Shalom, G.; Bevilacqua, B.; and Maron, H

Pith/arXiv arXiv 2006
[2021]

Chen, D.; O’Bray, L.; and Borgwardt, K

How attentive are graphattentionnetworks?arXivpreprintarXiv:2105.14491. Chen, D.; O’Bray, L.; and Borgwardt, K

Pith/arXiv arXiv
[2022]

Ding, Y.; Orvieto, A.; He, B.; and Hofmann, T

Understand- ing convolution on graphs via energies.arXiv preprint arXiv:2206.10991. Ding, Y.; Orvieto, A.; He, B.; and Hofmann, T

Pith/arXiv arXiv
[2023]

Bodnar, C.; Di Giovanni, F.; Chamberlain, B.; Lio, P.; and Bronstein, M

Specformer: Spectral graph neural networks meet transformers.arXiv preprint arXiv:2303.01028. Bodnar, C.; Di Giovanni, F.; Chamberlain, B.; Lio, P.; and Bronstein, M

Pith/arXiv arXiv
[2024]

Barbero, F.; Bodnar, C.; de Ocáriz Borde, H

Subgraphormer: Unifying Subgraph GNNs and Graph Transformers via Graph Products.arXiv preprint arXiv:2402.08450. Barbero, F.; Bodnar, C.; de Ocáriz Borde, H. S.; Bronstein, M.;Veličković,P.;andLiò,P.2022a. Sheafneuralnetworks with connection laplacians. InTopological, Algebraic and Geometric Learning Workshops 2022, 28–36. PMLR. Barbero, F.; Bodnar, C.; d...

Pith/arXiv arXiv 2022
[2025]

Freitas, S.; and Dong, Y

Sheaves Reloaded: A Direc- tional Awakening.arXiv preprint arXiv:2506.02842. Freitas, S.; and Dong, Y

Pith/arXiv arXiv

[1] [4]

Brody, S.; Alon, U.; and Yahav, E

Heterogeneous sheaf neural networks.arXiv preprint arXiv:2409.08036. Brody, S.; Alon, U.; and Yahav, E

Pith/arXiv arXiv

[2] [5]

We do not use mixed precision, and all reported numbers are produced in float32. Component Version Python 3.10.20 PyTorch 2.2.0 (+cu118) CUDA toolkit 11.8 PyTorch Geometric 2.3.1 pyg-lib0.4.0 torch-scatter2.1.2 torch-sparse0.6.18 torch-cluster1.6.3 torchmetrics1.9.0 GPU NVIDIA A100 40–80GB Table 5: Software and hardware environment. The com- plete depende...

2025

[3] [6]

Di Giovanni, F.; Rowbottom, J.; Chamberlain, B

Topology-Informed Graph Transformer.arXiv preprint arXiv:2402.02005. Di Giovanni, F.; Rowbottom, J.; Chamberlain, B. P.; Markovich, T.; and Bronstein, M. M

arXiv

[4] [8]

Errica, F.; Podda, M.; Bacciu, D.; and Micheli, A

Long range graph benchmark.arXiv preprint arXiv:2206.08164. Errica, F.; Podda, M.; Bacciu, D.; and Micheli, A

Pith/arXiv arXiv

[5] [11]

arXiv preprint arXiv:2012.06333

Sheaf neural networks. arXiv preprint arXiv:2012.06333. Hansen, J.; and Ghrist, R

Pith/arXiv arXiv 2012

[6] [12]

InInternational conference on machine learning, 12724–12745

A generalization of vit/mlp-mixer to graphs. InInternational conference on machine learning, 12724–12745. PMLR. Hu, W.; Fey, M.; Zitnik, M.; Dong, Y.; Ren, H.; Liu, B.; Catasta,M.;andLeskovec,J.2020. Opengraphbenchmark: Datasetsformachinelearningongraphs.Advancesinneural information processing systems, 33: 22118–22133. Huang, Y.; Miao, S.; and Li, P

2020

[7] [13]

Hussain, M

What Can We Learn from State Space Models for Machine Learning on Graphs? arXiv preprint arXiv:2406.05815. Hussain, M. S.; Zaki, M. J.; and Subramanian, D

Pith/arXiv arXiv

[8] [14]

InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 655–665

Globalself-attentionasareplacementforgraphconvolution. InProceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, 655–665. Kipf,T.N.;andWelling,M.2017.Semi-SupervisedClassifi- cationwithGraphConvolutionalNetworks. InInternational Conference on Learning Representations. Li, Q.; Han, Z.; and Wu, X.-M

2017

[9] [15]

Luo, Y.; Shi, L.; and Wu, X.-M

Graph Ex- ternal Attention Enhanced Transformer.arXiv preprint arXiv:2405.21061. Luo, Y.; Shi, L.; and Wu, X.-M

Pith/arXiv arXiv

[10] [16]

Ma, L.; Lin, C.; Lim, D.; Romero-Soriano, A.; Dokania, P

Can classic gnns be strong baselines for graph-level tasks? simple architectures meet excellence.arXiv preprint arXiv:2502.09263. Ma, L.; Lin, C.; Lim, D.; Romero-Soriano, A.; Dokania, P. K.; Coates, M.; Torr, P.; and Lim, S.-N

arXiv

[11] [17]

arXiv preprint arXiv:2305.17589

Graph Inductive Biases in Transformers without Message Passing. arXiv preprint arXiv:2305.17589. Rampášek, L.; Galkin, M.; Dwivedi, V. P.; Luu, A. T.; Wolf, G.; and Beaini, D

Pith/arXiv arXiv

[12] [18]

Ribeiro, A.; Tenório, A

Recipe for a General, Powerful, Scalable Graph Transformer.arXiv preprint arXiv:2205.12454. Ribeiro, A.; Tenório, A. L.; Belieni, J.; Souza, A. H.; and Mesquita, D

Pith/arXiv arXiv

[13] [19]

arXiv preprint arXiv:2507.00647

Cooperative Sheaf Neural Networks. arXiv preprint arXiv:2507.00647. Sancak, K.; Hua, Z.; Fang, J.; Xie, Y.; Malevich, A.; Long, B.;Balin,M.F.;andÇatalyürek,Ü.V.2024. AScalableand Effective Alternative to Graph Transformers.arXiv preprint arXiv:2406.12059. Shirzad, H.; Velingker, A.; Venkatachalam, B.; Sutherland, D.J.;andSinop,A.K.2023. Exphormer:Sparsetr...

arXiv 2024

[14] [20]

Traversa,P.;FerrazdeArruda,G.;Vazquez,A.;andMoreno, Y.2023.Robustnessandcomplexityofdirectedandweighted metabolic hypergraphs

Where did the gap go? reassessing the long-range graph benchmark.arXiv preprint arXiv:2309.00367. Traversa,P.;FerrazdeArruda,G.;Vazquez,A.;andMoreno, Y.2023.Robustnessandcomplexityofdirectedandweighted metabolic hypergraphs. Veličković,P.;Cucurull,G.;Casanova,A.;Romero,A.;Liò, P.; and Bengio, Y

Pith/arXiv arXiv 2023

[15] [21]

Wu,S.;Sun,F.;Zhang,W.;Xie,X.;andCui,B.2022

Graph- mamba: Towards long-range graph sequence modeling with selective state spaces.arXiv preprint arXiv:2402.00789. Wu,S.;Sun,F.;Zhang,W.;Xie,X.;andCui,B.2022. Graph neural networks in recommender systems: a survey.ACM computing surveys, 55(5): 1–37. 8 Experimental Setup and Reproducibility 8.1 Code The code associated with this work will be released up...

Pith/arXiv arXiv 2022

[16] [2019]

Fiorini, S.; Aktas, H.; Duta, I.; Coniglio, S.; Morerio, P.; Del Bue, A.; and Liò, P

A fair comparison of graph neural networks for graph classifi- cation.arXiv preprint arXiv:1912.09893. Fiorini, S.; Aktas, H.; Duta, I.; Coniglio, S.; Morerio, P.; Del Bue, A.; and Liò, P

Pith/arXiv arXiv 1912

[17] [2020]

Bar-Shalom, G.; Bevilacqua, B.; and Maron, H

On the bottleneck of graph neuralnetworksanditspracticalimplications.arXivpreprint arXiv:2006.05205. Bar-Shalom, G.; Bevilacqua, B.; and Maron, H

Pith/arXiv arXiv 2006

[18] [2021]

Chen, D.; O’Bray, L.; and Borgwardt, K

How attentive are graphattentionnetworks?arXivpreprintarXiv:2105.14491. Chen, D.; O’Bray, L.; and Borgwardt, K

Pith/arXiv arXiv

[19] [2022]

Ding, Y.; Orvieto, A.; He, B.; and Hofmann, T

Understand- ing convolution on graphs via energies.arXiv preprint arXiv:2206.10991. Ding, Y.; Orvieto, A.; He, B.; and Hofmann, T

Pith/arXiv arXiv

[20] [2023]

Bodnar, C.; Di Giovanni, F.; Chamberlain, B.; Lio, P.; and Bronstein, M

Specformer: Spectral graph neural networks meet transformers.arXiv preprint arXiv:2303.01028. Bodnar, C.; Di Giovanni, F.; Chamberlain, B.; Lio, P.; and Bronstein, M

Pith/arXiv arXiv

[21] [2024]

Barbero, F.; Bodnar, C.; de Ocáriz Borde, H

Subgraphormer: Unifying Subgraph GNNs and Graph Transformers via Graph Products.arXiv preprint arXiv:2402.08450. Barbero, F.; Bodnar, C.; de Ocáriz Borde, H. S.; Bronstein, M.;Veličković,P.;andLiò,P.2022a. Sheafneuralnetworks with connection laplacians. InTopological, Algebraic and Geometric Learning Workshops 2022, 28–36. PMLR. Barbero, F.; Bodnar, C.; d...

Pith/arXiv arXiv 2022

[22] [2025]

Freitas, S.; and Dong, Y

Sheaves Reloaded: A Direc- tional Awakening.arXiv preprint arXiv:2506.02842. Freitas, S.; and Dong, Y

Pith/arXiv arXiv