OpenGT: A Comprehensive Benchmark For Graph Transformers

Jiachen Tang; Jiajun Bu; Jiawei Chen; Sheng Zhou; Sirui Chen; Zhonghao Wang

REVIEW 3 major objections 4 minor 31 references

OpenGT: A Comprehensive Benchmark For Graph Transformers

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read OpenGT re-implements sixteen graph models under one pipeline and claims graph transformers are most competitive on heterophilous graphs, while local-only attention fails on sparse graphs and degree-based positional encodings offer cheap…

desk verdict A usable unified benchmark library, but the flagship positional-encoding experiment is undermined by a likely degenerate GESP encoding. read the letter →

arxiv 2506.04765 v1 pith:WW77E7T2 submitted 2025-06-05 cs.LG

Jiachen Tang , Zhonghao Wang , Sirui Chen , Sheng Zhou , Jiawei Chen , Jiajun Bu This is my paper

classification cs.LG

keywords graphtransformersbenchmarkneuralnetworkspositionalencodingheterophilyattentionmechanismsscalabilityrepresentationlearning

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

OpenGT is a bid to make graph-transformer research comparable: the authors re-implement sixteen graph transformer and GNN models in one modular pipeline and evaluate them on node- and graph-level datasets that vary in homophily, sparsity, and scale. The paper aims to establish when and why graph transformers beat message-passing GNNs, and which design choices actually matter. Its headline observations are that graph transformers are most competitive on heterophilous graphs, that purely local attention hurts on sparse graphs, that partition-based attention balances accuracy and cost, that degree-based positional encodings help on dense graphs, and that computing fancy positional encodings can cost more than training. If these observations hold, OpenGT gives the field a standard evaluation suite and a shortlist of design principles for future graph transformers.

What carries the argument

The load-bearing object is the OpenGT library itself: sixteen models re-implemented inside one modular graph-learning framework so that input encoders, attention layers, pooling, and heads are interchangeable, with shared splits and hyperparameter search. The controlled variable that drives several observations is the positional-encoding suite (degree-only GE, degree-plus-shortest-path GESP, Laplacian eigenvectors, equivariant-stable Laplacian, random-walk, and Weisfeiler-Lehman-style tokens) inserted unchanged into GraphGPS, DIFFormer, and SGFormer. The attention taxonomy (local, global, hybrid) and partition-based models supply the scaffolding for the efficiency claims.

What would settle it

Compute the GESP encoding for a small graph where two nodes have the same degree but very different distance profiles (for example, a path and a star) and inspect the attention-bias entries: if the shortest-path component never changes any attention value, or if GE and GESP produce bit-identical encodings, the paper's Observation 5 collapses.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is a set of empirical regularities: graph transformers outperform vanilla GNNs most clearly when the graph is heterophilous, models built for node-level prediction do not transfer reliably to graph-level tasks, local-attention-only models degrade on small sparse heterophilous graphs, graph-partitioned attention achieves an attractive accuracy/efficiency trade-off, degree-based Graphormer-style positional encodings are effective on dense graphs, and positional-encoding preprocessing (random walks, spectral decomposition, all-pairs shortest paths) can exceed training time on large graphs. The claimed mechanism is that global or hybrid attention gives graph transformers flexible receptive fields that matter when neighborhood information is unreliable, while the right cheap structural encoding supplies enough signal in dense graphs.

Load-bearing premise

The benchmark's conclusions stand or fall with the correctness of its re-implementations, and in particular with GESP actually incorporating shortest-path information rather than reducing to the degree-only GE encoding, since the paper reports identical results for the two.

Editorial extensions

If this is right

A new graph transformer can be tested against sixteen existing models under identical splits, compute budgets, and hyperparameter search, making reported gains interpretable.
Degree-based positional encodings become the default cheap first try for dense graphs, reserving spectral and shortest-path encodings for settings where they actually pay.
Local-attention-only designs need a global or hybrid component when deployed on sparse heterophilous graphs.
Partition-based attention is a viable route to scale graph transformers without sacrificing the accuracy of full attention.
Graph foundation-model efforts should treat node-level and graph-level objectives as distinct design targets, not interchangeable transfer tasks.

Reading between the lines

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

Beyond the paper, if the heterophily advantage is real, heterophilous datasets should become a standard slice of every graph-transformer evaluation, not a stress test.
Beyond the paper, the near-identical results reported for GE and GESP across all models and datasets suggest the shortest-path component may not be changing the model, so a direct check of GESP's attention bias would tell whether the degree-encoding finding extends to distance-aware encodings.
Beyond the paper, positional-encoding cost could be reduced adaptively: estimate graph density first and skip expensive encodings on dense graphs where degree information already helps.
Beyond the paper, future benchmarks could add a controlled ablation that isolates attention scope from model scale, since attention type and parameter count are confounded in the current comparison.

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.

Desk Editor's Note

read the letter

Two things to know about this paper. First, the artifact is real: a unified GraphGym-based library with 16 re-implemented GTs and GNNs on 14 node- and graph-level datasets, with standardized splits, hyperparameter search, and timing. That is genuinely useful for a fragmented field, and the authors mostly deliver it. Second, the paper's most distinctive analytical claim—that degree-based position encodings are as good as the full Graphormer encoding on dense graphs—is not supported by the tables. Look at Table 4: DIFFormer+GE and DIFFormer+GESP are identical to the printed standard error on all eight datasets. GPS+GE and GPS+GESP are identical on seven and differ by 0.0008 on Chameleon; SGFormer matches on four and stays within 0.006 elsewhere. That pattern is what you expect if the shortest-path component of GESP is never actually called and both columns run the same degree-only encoding. Observation 5's claim that GE is comparable to GESP becomes vacuous, and the benchmark's promise of including Graphormer's full encoding is broken. This is not a cosmetic issue; it is an implementation-correctness bug that needs to be fixed or the observation withdrawn.

Other soft spots are smaller. The efficiency claims in Observation 4 rest on two datasets (Chameleon and Citeseer), which is thin. Preprocessing times in Observation 6 are reported anecdotally ("about 10 minutes", "one hour") without a measurement protocol. Several conclusions are drawn from OOM-truncated model subsets, which biases the comparisons. There are no statistical tests, just mean±std over three seeds. And the exact data split definitions are not specified, which matters for a benchmark whose selling point is reproducibility. All of these are fixable.

What the paper does well is the engineering: a modular design, consistent training pipeline, and a broad set of models and datasets. The observations about heterophily and local attention are broadly consistent with prior work, so they are plausible even if not deeply novel. The benchmark, once cleaned up, could become a standard evaluation suite. Who benefits: graph transformer researchers who want a common harness and a library to build on. It deserves a serious referee, but the referee should ask the authors to verify the GESP implementation before accepting the positional-encoding conclusions. I would not cite it in its current form; I would revisit once the code is fixed.

Referee Report

3 major / 4 minor

Summary. The paper presents OpenGT, a benchmark library and standardized evaluation suite for graph transformers (GTs) and graph neural networks (GNNs). It re-implements 16 models in a common GraphGym-based pipeline, evaluates them on nine node-level and five graph-level datasets, and reports performance, time efficiency, and positional-encoding ablation results. From these experiments the paper draws six observations, including that GTs are more competitive on heterophilous graphs, that local attention is limited on sparse graphs, that graph partitioning improves efficiency, that degree-based positional encodings are effective on dense graphs, and that positional-encoding preprocessing is expensive on large graphs. The authors release code and configurations to support reproducibility.

Significance. If the reported observations are verified, OpenGT would be a useful community resource: it standardizes a fragmented evaluation landscape, covers both node- and graph-level tasks, and provides a modular implementation platform. The paper's main strengths are its open-source code release, the breadth of models and datasets, and the use of a consistent training pipeline with repeated seeds. However, the positional-encoding analysis is central to the paper's contribution, and the evidence for the key GESP-related observation is currently undermined by suspiciously identical results in Table 4. The benchmark's fairness and reproducibility claims depend on resolving this implementation-level concern.

major comments (3)

[Appendix A.2, Table 4; Section 4.4, Observation 5] The reported GE and GESP results are identical or near-identical across all testbeds: DIFFormer+GE and DIFFormer+GESP agree to four decimals on every node-level dataset (e.g., Cora 0.7467±0.0046, Chameleon 0.5446±0.0010), GPS+GE and GPS+GESP agree on seven of eight datasets, and SGFormer+GE/GESP agree within 0.006 everywhere. Exact agreement across independent random seeds is what one would expect if the same degree-only encoding were called twice rather than two distinct encodings. Since Observation 5 states that GE 'achieves performance comparable to GESP,' the comparison is vacuous unless GESP actually incorporates the shortest-path component of Graphormer's encoding. I ask the authors to verify the implementation with unit tests on small graphs with known shortest paths, provide the precise definition of GE and GESP, and re-run the affected experiments. Until this is resolved, the PE guidance in Sections 4.4 and 5 is unsupported.
[Section 4.4] The paper never defines GE or GESP formally: there is no equation, no description of the input tensors, and no statement of how either encoding is injected into DIFFormer, GPS, or SGFormer. Without these definitions, readers cannot assess whether the implementation matches Graphormer's original bias encoding, whether the shortest-path term is actually computed, or whether the integration mechanism is identical across the three testbed models. I request explicit formulas for GE and GESP and a precise description of the integration scheme, both in the main text or appendix and in the code documentation.
[Section 4.4, Observation 6] Observation 6 claims that preprocessing positional encodings is expensive on large graphs, giving approximate times of ~10 minutes for random-walk statistics, ~1 hour for spectral decomposition, and up to ~2 hours for all-pairs shortest paths on PubMed. No measurement protocol, timing table, or script pointer is provided, and the time-efficiency tables in Appendix A.3 exclude preprocessing by design. Because this observation is explicitly used to motivate a future research direction, the authors should report the preprocessing-time measurements, the hardware and version details, and the protocol (e.g., number of runs and seeds) just as they do for training-time measurements.

minor comments (4)

[Abstract and Section 4.1] There are several typos and grammatical issues, including 'this highlights the need' in the abstract, 'compariable' in Observation 1, 'evalauted' in the Table 1 caption, 'raise an expection' in Observation 2, and 'Liscence' in Appendix C. A careful proofread is needed.
[Figures 2 and 3] The manuscript references Figures 2 and 3 for the performance overview, but the actual figures are not visible in the text provided. Please ensure the final version contains the figures and that the captions explain the meaning of empty bars and OOM markers.
[Section 3.1] The claim of a 'fair comparison' rests on a shared hyperparameter grid for all models. Models with specialized components, such as GRIT or Exphormer, may require different learning-rate schedules or hidden dimensions; the paper should state whether any model-specific hyperparameters from the original papers were retained and how the grid was chosen to avoid systematically disadvantaging certain architectures.
[Appendix B, Table 7] The hyperparameter search space lists only learning rate, weight decay, number of layers, number of heads, and dropout. It does not mention training epochs, patience, or the criterion for selecting the best configuration (e.g., best validation metric vs. final epoch). Adding these details would strengthen reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: OpenGT is an empirical benchmark whose observations are not entailed by its own inputs; the identical GE/GESP numbers in Table 4 are a correctness concern, not a circularity.

full rationale

This paper is an empirical benchmark study rather than a derivation chain. It does not fit parameters and then rename those fits as predictions, and its observations are not logically entailed by its experimental setup. The categorization of models by attention scope and positional-encoding families is based on published architectural mechanisms, not on the performance outcomes being explained, so no self-definitional loop is present. Observation 5 ('GE achieves performance comparable to GESP') depends on GESP being implemented with shortest-path information, but even if GESP and GE were identical in the released code, that would be an implementation-correctness or validity issue, not circular reasoning: the paper's stated definition of GESP explicitly includes shortest-path information, and the reported equivalence is a factual claim about experimental results rather than a conclusion defined in terms of its own premises. There are no load-bearing self-citations, no imported uniqueness theorems, and no ansatz smuggled in via citation. The identical GE/GESP cells in Table 4 (e.g., DIFFormer+GE and DIFFormer+GESP both 0.5446±0.0010 on Chameleon, and matching values on all other datasets) merit scrutiny as a potential implementation bug, but that scrutiny belongs to correctness review, not circularity analysis. Accordingly, the appropriate finding is no significant circularity.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The benchmark is an empirical infrastructure contribution, so no mathematical axioms are needed. Its conclusions rest on implementation faithfulness, tuning adequacy, and correct positional encoding implementation. The GE/GESP identity is the most concrete threat.

free parameters (3)

shared hyperparameter grid
Learning rate, weight decay, layers, heads, dropout are grid-searched over one space for all models (Appendix B), which can favor some models over others.
data split scheme
The paper states same splits are used but does not specify split ratios or procedure, making the numbers hard to exactly reproduce.
batch size per graph-level dataset
Batch sizes are fixed within a dataset but vary between datasets; the choice is not justified in the paper.

assumptions (5)

domain assumption The re-implementations of all 16 models are faithful to the original papers.
The benchmark's comparisons rest on this; no external verification or diff against reference implementations is provided in the paper.
domain assumption The shared hyperparameter search space is adequate for every model.
Appendix B defines one grid for all models; some models may require different tuning to avoid underperformance.
ad hoc to paper GESP correctly incorporates shortest-path information and is distinct from GE.
Table 4 shows GE and GESP produce identical or near-identical results across models and datasets, suggesting the shortest-path component is inactive.
domain assumption Models that OOM are excluded without biasing the conclusions.
Tables 2 and 3 contain many OOM entries; the reported observations use only the subset of models that fit in 24GB memory.
domain assumption The selected datasets adequately represent graph learning scenarios.
The paper claims coverage of scale, homophily, and sparsity, but the classification of 'dense' and 'sparse' is qualitative and not quantified.

how reviews work

0 comments

Cite this review

Pith. "Pith review of OpenGT: A Comprehensive Benchmark For Graph Transformers." pith.science (2026). https://pith.science/paper/WW77E7T2

@misc{pith2026250604765,
  author       = {Pith},
  title        = {Pith review of: OpenGT: A Comprehensive Benchmark For Graph Transformers},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WW77E7T2}},
  note         = {Machine review of arXiv:2506.04765}
}

read the original abstract

Graph Transformers (GTs) have recently demonstrated remarkable performance across diverse domains. By leveraging attention mechanisms, GTs are capable of modeling long-range dependencies and complex structural relationships beyond local neighborhoods. However, their applicable scenarios are still underexplored, this highlights the need to identify when and why they excel. Furthermore, unlike GNNs, which predominantly rely on message-passing mechanisms, GTs exhibit a diverse design space in areas such as positional encoding, attention mechanisms, and graph-specific adaptations. Yet, it remains unclear which of these design choices are truly effective and under what conditions. As a result, the community currently lacks a comprehensive benchmark and library to promote a deeper understanding and further development of GTs. To address this gap, this paper introduces OpenGT, a comprehensive benchmark for Graph Transformers. OpenGT enables fair comparisons and multidimensional analysis by establishing standardized experimental settings and incorporating a broad selection of state-of-the-art GNNs and GTs. Our benchmark evaluates GTs from multiple perspectives, encompassing diverse tasks and datasets with varying properties. Through extensive experiments, our benchmark has uncovered several critical insights, including the difficulty of transferring models across task levels, the limitations of local attention, the efficiency trade-offs in several models, the application scenarios of specific positional encodings, and the preprocessing overhead of some positional encodings. We aspire for this work to establish a foundation for future graph transformer research emphasizing fairness, reproducibility, and generalizability. We have developed an easy-to-use library OpenGT for training and evaluating existing GTs. The benchmark code is available at https://github.com/eaglelab-zju/OpenGT.

Figures

Figures reproduced from arXiv: 2506.04765 by the authors.

**Figure 2.** The performance overview on node-level tasks. The empty bar denotes that the model [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗

**Figure 3.** The performance overview on graph-level tasks. The metric used by each dataset is written [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Time consumption and Test accuracy of different methods on Chameleon and Citeseer [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

**Figure 5.** Figure 5: Accuracy results different models integrated with different positional encodings. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

31 extracted references · 8 canonical work pages

[1]

Specformer: Spectral graph neural networks meet transformers.arXiv preprint arXiv:2303.01028, 2023

Deyu Bo, Chuan Shi, Lele Wang, and Renjie Liao. Specformer: Spectral graph neural networks meet transformers.arXiv preprint arXiv:2303.01028, 2023

arXiv 2023
[2]

A generalization of transformer networks to graphs.CoRR, abs/2012.09699, 2020

Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs.CoRR, abs/2012.09699, 2020

arXiv 2012
[3]

Long range graph benchmark.Advances in Neural Information Processing Systems, 35:22326–22340, 2022

Vijay Prakash Dwivedi, Ladislav Rampášek, Michael Galkin, Ali Parviz, Guy Wolf, Anh Tuan Luu, and Dominique Beaini. Long range graph benchmark.Advances in Neural Information Processing Systems, 35:22326–22340, 2022

2022
[4]

Predict then propagate: Graph neural networks meet personalized pagerank.arXiv preprint arXiv:1810.05997, 2018

Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate: Graph neural networks meet personalized pagerank.arXiv preprint arXiv:1810.05997, 2018

arXiv 2018
[5]

Benchmarking positional encodings for gnns and graph transformers.arXiv preprint arXiv:2411.12732, 2024

Florian Grötschla, Jiaqing Xie, and Roger Wattenhofer. Benchmarking positional encodings for gnns and graph transformers.arXiv preprint arXiv:2411.12732, 2024

work page arXiv 2024
[6]

A generaliza- tion of vit/mlp-mixer to graphs

Xiaoxin He, Bryan Hooi, Thomas Laurent, Adam Perold, Yann LeCun, and Xavier Bresson. A generaliza- tion of vit/mlp-mixer to graphs. InInternational conference on machine learning, pages 12724–12745. PMLR, 2023

work page 2023
[7]

Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems, 33:22118–22133, 2020

Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems, 33:22118–22133, 2020

2020
[8]

Zinc: a free tool to discover chemistry for biology.Journal of chemical information and modeling, 52(7):1757–1768, 2012

John J Irwin, Teague Sterling, Michael M Mysinger, Erin S Bolstad, and Ryan G Coleman. Zinc: a free tool to discover chemistry for biology.Journal of chemical information and modeling, 52(7):1757–1768, 2012

work page 2012

Show all 31 references

[9]

The impact of positional encoding on length generalization in transformers

Amirhossein Kazemnejad, Inkit Padhi, Karthikeyan Natesan Ramamurthy, Payel Das, and Siva Reddy. The impact of positional encoding on length generalization in transformers. In A. Oh, T. Naumann, A. Globerson, K. Saenko, M. Hardt, and S. Levine, editors,Advances in Neural Inform...

2023
[10]

Semi-supervised classification with graph convolutional networks

Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional networks. arXiv preprint arXiv:1609.02907, 2016

2016 arXiv
[11]

Rethinking graph transformers with spectral attention.Advances in Neural Information Processing Systems, 34:21618– 21629, 2021

Devin Kreuzer, Dominique Beaini, Will Hamilton, Vincent Létourneau, and Prudencio Tossou. Rethinking graph transformers with spectral attention.Advances in Neural Information Processing Systems, 34:21618– 21629, 2021

2021
[12]

Towards graph foundation models: A survey and beyond.arXiv preprint arXiv:2310.11829, 2023

Jiawei Liu, Cheng Yang, Zhiyuan Lu, Junze Chen, Yibo Li, Mengmei Zhang, Ting Bai, Yuan Fang, Lichao Sun, Philip S Yu, et al. Towards graph foundation models: A survey and beyond.arXiv preprint arXiv:2310.11829, 2023

2023 arXiv
[13]

Graph inductive biases in transformers without message passing

Liheng Ma, Chen Lin, Derek Lim, Adriana Romero-Soriano, Puneet K Dokania, Mark Coates, Philip Torr, and Ser-Nam Lim. Graph inductive biases in transformers without message passing. InInternational Conference on Machine Learning, pages 23321–23337. PMLR, 2023

2023
[14]

Geom-gcn: Geometric graph convolutional networks.arXiv preprint arXiv:2002.05287, 2020

Hongbin Pei, Bingzhe Wei, Kevin Chen-Chuan Chang, Yu Lei, and Bo Yang. Geom-gcn: Geometric graph convolutional networks.arXiv preprint arXiv:2002.05287, 2020

2002 arXiv
[15]

Recipe for a general, powerful, scalable graph transformer.Advances in Neural Information Processing Systems, 35:14501–14515, 2022

Ladislav Rampášek, Michael Galkin, Vijay Prakash Dwivedi, Anh Tuan Luu, Guy Wolf, and Dominique Beaini. Recipe for a general, powerful, scalable graph transformer.Advances in Neural Information Processing Systems, 35:14501–14515, 2022

2022
[16]

Gemsec: Graph embedding with self clustering

Benedek Rozemberczki, Ryan Davies, Rik Sarkar, and Charles Sutton. Gemsec: Graph embedding with self clustering. InProceedings of the 2019 IEEE/ACM international conference on advances in social networks analysis and mining, pages 65–72, 2019

2019
[17]

Collective classification in network data.AI magazine, 29(3):93–93, 2008

Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. Collective classification in network data.AI magazine, 29(3):93–93, 2008

2008
[18]

Pitfalls of graph neural network evaluation.arXiv preprint arXiv:1811.05868, 2018

Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. Pitfalls of graph neural network evaluation.arXiv preprint arXiv:1811.05868, 2018

2018 arXiv
[19]

Graph transformers: A survey.arXiv preprint arXiv:2407.09777, 2024

Ahsan Shehzad, Feng Xia, Shagufta Abid, Ciyuan Peng, Shuo Yu, Dongyu Zhang, and Karin Verspoor. Graph transformers: A survey.arXiv preprint arXiv:2407.09777, 2024. 10

2024
[20]

Exphormer: Sparse transformers for graphs

Hamed Shirzad, Ameya Velingker, Balaji Venkatachalam, Danica J Sutherland, and Ali Kemal Sinop. Exphormer: Sparse transformers for graphs. InInternational Conference on Machine Learning, pages 31613–31632. PMLR, 2023

2023
[21]

Attention is all you need.Advances in neural information processing systems, 30, 2017

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

2017
[22]

Graph attention networks.arXiv preprint arXiv:1710.10903, 2017

Petar Veliˇckovi´c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks.arXiv preprint arXiv:1710.10903, 2017

2017 arXiv
[23]

Equivariant and stable positional encoding for more powerful graph neural networks.arXiv preprint arXiv:2203.00199, 2022

Haorui Wang, Haoteng Yin, Muhan Zhang, and Pan Li. Equivariant and stable positional encoding for more powerful graph neural networks.arXiv preprint arXiv:2203.00199, 2022

2022 arXiv
[24]

Graph triple attention network: A decoupled perspective.arXiv preprint arXiv:2408.07654, 2024

Xiaotang Wang, Yun Zhu, Haizhou Shi, Yongchao Liu, and Chuntao Hong. Graph triple attention network: A decoupled perspective.arXiv preprint arXiv:2408.07654, 2024

2024 arXiv
[25]

Difformer: Scalable (graph) transformers induced by energy constrained diffusion.arXiv preprint arXiv:2301.09474, 2023

Qitian Wu, Chenxiao Yang, Wentao Zhao, Yixuan He, David Wipf, and Junchi Yan. Difformer: Scalable (graph) transformers induced by energy constrained diffusion.arXiv preprint arXiv:2301.09474, 2023

2023 arXiv
[26]

Nodeformer: A scalable graph structure learning transformer for node classification.Advances in Neural Information Processing Systems, 35:27387–27401, 2022

Qitian Wu, Wentao Zhao, Zenan Li, David P Wipf, and Junchi Yan. Nodeformer: A scalable graph structure learning transformer for node classification.Advances in Neural Information Processing Systems, 35:27387–27401, 2022

2022
[27]

Sgformer: Simplifying and empowering transformers for large-graph representations.Advances in Neural Information Processing Systems, 36:64753–64773, 2023

Qitian Wu, Wentao Zhao, Chenxiao Yang, Hengrui Zhang, Fan Nie, Haitian Jiang, Yatao Bian, and Junchi Yan. Sgformer: Simplifying and empowering transformers for large-graph representations.Advances in Neural Information Processing Systems, 36:64753–64773, 2023

2023
[28]

Less is more: on the over-globalizing problem in graph transformers.arXiv preprint arXiv:2405.01102, 2024

Yujie Xing, Xiao Wang, Yibo Li, Hai Huang, and Chuan Shi. Less is more: on the over-globalizing problem in graph transformers.arXiv preprint arXiv:2405.01102, 2024

2024 arXiv
[29]

Do transformers really perform badly for graph representation?Advances in neural information processing systems, 34:28877–28888, 2021

Chengxuan Ying, Tianle Cai, Shengjie Luo, Shuxin Zheng, Guolin Ke, Di He, Yanming Shen, and Tie-Yan Liu. Do transformers really perform badly for graph representation?Advances in neural information processing systems, 34:28877–28888, 2021

2021
[30]

Design space for graph neural networks.Advances in Neural Information Processing Systems, 33:17009–17021, 2020

Jiaxuan You, Zhitao Ying, and Jure Leskovec. Design space for graph neural networks.Advances in Neural Information Processing Systems, 33:17009–17021, 2020

2020
[31]

Rethinking positional encoding.arXiv preprint arXiv:2107.02561, 2021

Jianqiao Zheng, Sameera Ramasinghe, and Simon Lucey. Rethinking positional encoding.arXiv preprint arXiv:2107.02561, 2021. 11 A Additional Results A.1 Experiment results for different models In section 4, we investigated the performance of GTs and GNNs on different datasets. T...

2021 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

Specformer: Spectral graph neural networks meet transformers.arXiv preprint arXiv:2303.01028, 2023

Deyu Bo, Chuan Shi, Lele Wang, and Renjie Liao. Specformer: Spectral graph neural networks meet transformers.arXiv preprint arXiv:2303.01028, 2023

arXiv 2023

[2] [2]

A generalization of transformer networks to graphs.CoRR, abs/2012.09699, 2020

Vijay Prakash Dwivedi and Xavier Bresson. A generalization of transformer networks to graphs.CoRR, abs/2012.09699, 2020

arXiv 2012

[3] [3]

Long range graph benchmark.Advances in Neural Information Processing Systems, 35:22326–22340, 2022

Vijay Prakash Dwivedi, Ladislav Rampášek, Michael Galkin, Ali Parviz, Guy Wolf, Anh Tuan Luu, and Dominique Beaini. Long range graph benchmark.Advances in Neural Information Processing Systems, 35:22326–22340, 2022

2022

[4] [4]

Predict then propagate: Graph neural networks meet personalized pagerank.arXiv preprint arXiv:1810.05997, 2018

Johannes Gasteiger, Aleksandar Bojchevski, and Stephan Günnemann. Predict then propagate: Graph neural networks meet personalized pagerank.arXiv preprint arXiv:1810.05997, 2018

arXiv 2018

[5] [5]

Benchmarking positional encodings for gnns and graph transformers.arXiv preprint arXiv:2411.12732, 2024

Florian Grötschla, Jiaqing Xie, and Roger Wattenhofer. Benchmarking positional encodings for gnns and graph transformers.arXiv preprint arXiv:2411.12732, 2024

work page arXiv 2024

[6] [6]

A generaliza- tion of vit/mlp-mixer to graphs

Xiaoxin He, Bryan Hooi, Thomas Laurent, Adam Perold, Yann LeCun, and Xavier Bresson. A generaliza- tion of vit/mlp-mixer to graphs. InInternational conference on machine learning, pages 12724–12745. PMLR, 2023

work page 2023

[7] [7]

Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems, 33:22118–22133, 2020

Weihua Hu, Matthias Fey, Marinka Zitnik, Yuxiao Dong, Hongyu Ren, Bowen Liu, Michele Catasta, and Jure Leskovec. Open graph benchmark: Datasets for machine learning on graphs.Advances in neural information processing systems, 33:22118–22133, 2020

2020

[8] [8]

Zinc: a free tool to discover chemistry for biology.Journal of chemical information and modeling, 52(7):1757–1768, 2012

John J Irwin, Teague Sterling, Michael M Mysinger, Erin S Bolstad, and Ryan G Coleman. Zinc: a free tool to discover chemistry for biology.Journal of chemical information and modeling, 52(7):1757–1768, 2012

work page 2012