Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Rethinking Tokenized Graph Transformers for Node Classification

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

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

Pith's one-line read SwapGT claims that random token swapping between neighboring token sets enlarges a tokenized graph Transformer's sampling space from the 1-hop to the (t+1)-hop neighborhood of a k-NN graph and yields the best node-classification accuracy…

desk verdict SwapGT is a simple, well-tested token-swapping augmentation for tokenized graph Transformers, but its "semantic relevance" explanation is not supported by the evidence. read the letter →

arxiv 2502.08101 v1 pith:T55RUJWN submitted 2025-02-12 cs.LG cs.AI

Jinsong Chen , Chenyang Li , GaiChao Li , John E. Hopcroft , Kun He This is my paper · ORCID
classification cs.LGcs.AI
keywords nodeclassificationgraphtransformerstokenizedtokenswappingk-nearestneighborcenteralignmentlossdataaugmentationheterophily
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

Existing tokenized graph Transformers generate each node's token sequence by selecting the top-k most similar nodes on a similarity graph—in effect, reading only the first-order neighborhood of that graph. The paper argues that this wastes most of the node-pair correlations that the similarity graph encodes, and proposes SwapGT, whose token swapping operation replaces each token with a random token drawn from that token's own top-k set. Repeating this t times is equivalent to sampling from the (t+1)-hop neighborhood of the k-NN graph, so the Transformer sees semantically related nodes that previous methods could never include as tokens. A center alignment loss then keeps representations from the multiple token sequences close. SwapGT reports the best mean accuracy on all eight datasets, under both dense and sparse training splits, with the largest margins in the sparse regime.

What carries the argument

The load-bearing object is the token swapping operation on the k-NN graph. For each token v_j in a node's token set, with probability p the algorithm replaces v_j by a uniformly random member of v_j's own token set; iterating t times means every slot in the sequence is reachable by a walk of length up to t+1 from the target node in the k-NN graph. This is the mechanism that turns a 1-hop sampler into a multi-hop sampler without retraining or changing the Transformer backbone, while the center alignment loss, which uses cosine similarity to the mean sequence representation, is the auxiliary mechanism that prevents the augmented sequences from drifting apart.

What would settle it

Compute the class purity of tokens after t swaps: for a target node i, measure the probability that a swapped-in node shares i's label, and compare it to the global label prior. If, as t grows from 0 to 4, this conditional probability converges to the global prior rather than staying noticeably above it, the claim that swapping exploits semantic relevance is falsified—the gains would come from augmentation alone, not from reaching semantically related nodes.

Watch

Extended reading notes

Core claim

The central claim is that token generation in tokenized GTs is a first-order neighbor selection on a k-NN graph, and that performing token swapping t times enlarges the sampling space to the (t+1)-hop neighborhood. Because the k-NN graph is built from cosine similarity in raw attribute space and in a personalized-PageRank-smoothed topological space, the nodes reached through swapping remain semantically related to the target rather than being noise. The resulting s+1 token sequences per node are encoded by a standard Transformer, read out by taking the first token's representation concatenated with the mean of the other sequence representations, and fused across the two views with a scalar weight. Training minimizes cross-entropy together with a center alignment loss that pulls the representations of the multiple sequences of the same node toward their average. On eight datasets spanning homophily levels from 0.24 to 0.83, SwapGT achieves the top mean accuracy under both dense (50/25/25) and sparse (2.5/2.5/95) splits.

Load-bearing premise

The load-bearing premise is that the k-NN graph built from cosine similarity in raw attribute space and in a personalized-PageRank-smoothed topological space captures which nodes are semantically relevant to the target, so a random swap through a neighbor's token set is more likely to add an informative node than noise.

Editorial extensions

If this is right

  • Tokenized GTs can be made to see beyond first-order neighbors without any change to the Transformer backbone: the swap operation is preprocessing on the token sets.
  • The largest gains occur under the 2.5%/2.5%/95% sparse split, so token swapping acts as a data augmentation that helps the Transformer generalize when labels are scarce.
  • The ablation against a single long sequence (2k tokens) and against random resampling from the enlarged 2k set shows that multiple diverse short token sequences are preferable to one longer sequence, and that swapping is better than blind resampling.
  • The center alignment loss improves accuracy on every dataset and helps most under sparse supervision, making it a reusable regularizer for any model that assigns several token sequences to one node.

Reading between the lines

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

  • [Editorial inference] The method's success should depend on how well the k-NN graph reflects label-relevant similarity; on graphs where feature similarity is a poor proxy for class, swapping could pull in off-class nodes, so the approach would need a similarity graph learned or filtered during training.
  • [Editorial inference] The (t+1)-hop equivalence is stated for the k-NN graph; a direct testable extension is to apply the same swap-and-align recipe to a GNN-based token sampler or to a k-NN graph built from learned embeddings, which would separate the swapping mechanism from the specific similarity measure.
  • [Editorial inference] Because the center alignment loss acts only within each view (attribute and topology separately) and the views are linearly fused, a natural extension is to add cross-view alignment so the two views constrain each other; the paper's reported gains leave open how much of the improvement is intra-view consistency versus the added diversity of sequences.
  • [Editorial inference] A cheaper variant could use one swapped sequence per node during training and average over swaps at inference, which the paper does not explore.
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 proposes SwapGT, a tokenized graph Transformer for node classification. SwapGT constructs initial token sequences from attribute and topology views using k-NN sampling, then applies a stochastic token swapping operation that replaces tokens with randomly drawn tokens from their own token sets, iterated t times. The authors claim this enlarges the sampling space from the 1-hop to the (t+1)-hop neighborhood of the k-NN graph and thereby produces more informative token sequences. A center alignment loss regularizes representations across s augmented sequences. Experiments on eight datasets under dense and sparse splits report that SwapGT outperforms eleven baselines, with ablations on the center loss, swapping versus random subsetting, and the hyperparameters t and s.

Significance. If the mechanism is as claimed, the paper would contribute a simple augmentation technique for tokenized graph Transformers, potentially improving performance on heterophilous graphs under sparse supervision. The idea of enlarging the candidate pool via graph-structure swaps is intuitive, and the reported gains are consistent across datasets. However, the paper's central interpretative claim—that the benefit comes from semantic relevance rather than generic token diversity—is not established by the current evidence, and the empirical support lacks significance testing and a clean control. The contributions are incremental but potentially useful.

major comments (5)
  1. [Section 4.2, Eq. (5), Algorithm 1] The claim that performing swapping t times 'is equal to enlarge the sampling space from 1-hop neighborhood to (t+1)-hop neighborhood' is only an upper bound on reachability, not an equality: with probability p in Algorithm 1 a token may remain in place, and the final set has only k tokens, so it is not the full (t+1)-hop neighborhood. More importantly, the uniform random replacement in Eq. (5) is an unweighted random step on the k-NN graph, so the hop-count argument does not imply that the newly reachable nodes are semantically relevant. The paper provides no direct evidence—such as label purity of swapped tokens, cosine-similarity distributions of the expanded candidate set, or attention analysis—that the enlarged pool contains more informative nodes rather than more noise. This unsupported interpretative claim appears in the abstract and in Section 4.2.
  2. [Section 5.5, Figure 5] The comparison with SwapGT-R does not isolate the proposed multi-hop mechanism. SwapGT-R samples k tokens from the top-2k set, so its candidate pool is fixed at size 2k, whereas SwapGT's swapping can reach nodes at distance up to t+1 in the k-NN graph, i.e., a much larger pool. The observed gains over SwapGT-R could therefore be explained by the larger candidate pool or by the specific stochastic walk, rather than by semantic relevance. A proper control would hold the candidate-pool size comparable (e.g., sampling from top-mk tokens for m matching the reachable set size) while removing the graph-structure dependence.
  3. [Sections 5.6 and 5.7, Figures 6 and 7] The paper reports accuracy as a function of the swapping times t and augmentation times s, but does not state whether these hyperparameters were selected on the validation set or on the test set. If the reported curves and the final results pick t and s based on test performance, the 'best on all datasets' claim is optimistically biased. The authors should specify the selection protocol and, if necessary, re-tune the hyperparameters using only training/validation data.
  4. [Tables 1 and 2] The claim that SwapGT is best on all eight datasets is not supported by any significance testing. Several gaps are within one standard deviation of the best baseline (e.g., Flickr dense: 87.56±0.61 vs PolyFormer 87.01±0.57; WikiCS sparse: 78.11±0.83 vs BM-GCN 77.39±0.37; Citeseer dense: 78.49±0.95 vs PolyFormer 78.03±0.86). Without paired tests across the ten runs, the differences cannot be assessed. The authors should report paired significance tests (e.g., Wilcoxon signed-rank or paired t-test) or quantify effect sizes.
  5. [Section 4.2, Eq. (5) vs Algorithm 1] There is an inconsistency between the formulation and the implementation: Eq. (5) states that every token is replaced (N'_i = {ζ(N_j) | v_j in N_i}), while Algorithm 1 replaces a token only if random(0,1) > p. Since the theoretical discussion relies on Eq. (5), the paper should clarify that Eq. (5) describes the p=1 case or otherwise align the mathematical formulation with the actual algorithm.
minor comments (6)
  1. [Section 2 heading] The heading 'Relation Work' should be 'Related Work'.
  2. [Table 3] The header 'H ↓' is ambiguous because lower H indicates stronger heterophily; please add a footnote or clarify the meaning of the arrow.
  3. [Section 5.3] There is a typo: 'GTs usually leg behind GNNs' should be 'lag behind'.
  4. [Section 4.3, Eq. (7)] The model input dimension in Eq. (7) is written as R^{(1+s)×k×d0}, which appears to omit the target node token whose presence is indicated in Eq. (6); please check the dimension consistency.
  5. [Appendix B.2] The statement 'SwapGT applies the uniform strategy for selecting node tokens' is unclear because the swapping operation uses uniform random sampling; please clarify what 'uniform strategy' refers to.
  6. [Reproducibility] No code or link to code is provided; releasing an implementation would improve reproducibility and help readers verify the empirical claims.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the hop-expansion claim follows directly from the definition of token swapping, and no fitted parameter is relabeled as a prediction.

full rationale

The paper's only derivation-like claim is the statement in Section 4.2 that t token-swapping operations enlarge the sampling space from the 1-hop to the (t+1)-hop neighborhood of the k-NN graph. This is a graph-theoretic consequence of Eq. 5: each replacement draws uniformly from the token set N_j of a 1-hop neighbor v_j, so the set of possible replacements after one swap is exactly the 2-hop neighborhood, and iterating t times gives the (t+1)-hop neighborhood. The claim is not used to define swapping; it is derived from the definition, so no self-definitional circularity is present. The 'semantic relevance' language adds an interpretive gloss, but the experimental gains are measured on held-out node classification accuracy against eleven external baselines, not predicted from the k-NN inputs. Hyperparameters such as k, alpha, t, s, and p are tuned per dataset, but no fitted parameter is renamed as a prediction. The paper cites the authors' prior NAGphormer and NTFormer work only as background and as baselines; no load-bearing uniqueness theorem or ansatz is imported from those papers. The SwapGT-R and SwapGT-L ablations are comparative controls rather than circular restatements; the fact that SwapGT-R may not isolate the multi-hop mechanism is an experimental-design concern, not circularity. Overall, the central derivation is self-contained, and the contribution stands or falls on empirical evidence.

Assumptions & free parameters 11 free parameters · 5 assumptions · 0 invented entities

The paper is an empirical method paper. The central performance claim depends on several free hyperparameters, at least four of which (p, lambda, K, L) are never specified, and on the domain assumption that random swapping in a cosine-similarity k-NN graph produces informative tokens. No new physical or formal entities are introduced.

free parameters (11)
  • k (number of tokens per sequence) = tuned in {4,6,8}; final value per dataset not reported
    Determines the size of token sets in Eq. 4 and thus the swapping graph.
  • alpha (view fusion weight) = tuned in {0.1,...,0.9}; final not reported
    Balances attribute and topology representations in Eq. 11.
  • p (swapping probability) = not reported
    Controls how often each token is swapped in Algorithm 1; never specified.
  • t (swapping times) = analyzed in {1,2,3,4}; final not reported
    Sets the effective hop range in Section 4.2; treated as a free hyperparameter.
  • s (augmentation times) = analyzed in {1..8}; final not reported
    Determines number of token sequences per node; central to the method.
  • lambda (center alignment loss weight) = not reported
    Weights Lca in Eq. 16; never specified or tuned in text.
  • K (PPR propagation steps) = not reported
    Number of propagation steps in X' = phi(Â, X, K), Section 4.1; never specified.
  • L (number of Transformer layers) = not reported
    Depth of encoder in Eqs. 8-9; never specified.
  • learning rate = grid {0.001,0.005,0.01}
    Standard training hyperparameter searched per dataset.
  • dropout = grid {0.3,0.5,0.7}
    Standard training hyperparameter searched per dataset.
  • hidden dimension = grid {256,512}
    Model width searched per dataset.
assumptions (5)
  • standard math Transformer multi-head self-attention and FFN are computed as in Eqs. 1-3.
    Background from Vaswani et al. 2017; not derived.
  • domain assumption Personalized PageRank propagation produces topology features that represent graph structure.
    Adopted from Klicpera et al. 2019, used in Section 4.1.
  • domain assumption Cosine similarity between node features reflects semantic relevance useful for node classification.
    Basis for constructing the k-NN token graph in Section 4.1.
  • ad hoc to paper Randomly replacing a token with a random token from that token's own set yields informative diverse sequences.
    Core innovation of SwapGT, Eq. 5; no proof that swapped tokens are more semantically relevant than originals.
  • domain assumption Representations from different token sequences are views of the same node and should be aligned in latent space.
    Motivates the center alignment loss, Eq. 14.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rethinking Tokenized Graph Transformers for Node Classification." pith.science (2026). https://pith.science/paper/T55RUJWN

@misc{pith2026250208101,
  author       = {Pith},
  title        = {Pith review of: Rethinking Tokenized Graph Transformers for Node Classification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/T55RUJWN}},
  note         = {Machine review of arXiv:2502.08101}
}
read the original abstract

Node tokenized graph Transformers (GTs) have shown promising performance in node classification. The generation of token sequences is the key module in existing tokenized GTs which transforms the input graph into token sequences, facilitating the node representation learning via Transformer. In this paper, we observe that the generations of token sequences in existing GTs only focus on the first-order neighbors on the constructed similarity graphs, which leads to the limited usage of nodes to generate diverse token sequences, further restricting the potential of tokenized GTs for node classification. To this end, we propose a new method termed SwapGT. SwapGT first introduces a novel token swapping operation based on the characteristics of token sequences that fully leverages the semantic relevance of nodes to generate more informative token sequences. Then, SwapGT leverages a Transformer-based backbone to learn node representations from the generated token sequences. Moreover, SwapGT develops a center alignment loss to constrain the representation learning from multiple token sequences, further enhancing the model performance. Extensive empirical results on various datasets showcase the superiority of SwapGT for node classification.

Figures

Figures reproduced from arXiv: 2502.08101 by the authors.

Figure 1
Figure 1. The toy example of token generation on the k-NN graph. Previous methods only focus on 1-hop neighborhood to construct a single token sequence. While our method can flexibly select tokens from multi-hop neighborhoods to generate diverse token sequences [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The overall framework of SwapGT. First, we generate the initial token sequences from both the attribute view and topology view. Then, we utilize the proposed token swapping operation to generate new token sequences for each target node. These generated token sequences are then fed into a Transformer-based backbone to learn node representations and generate predicted labels. Additionally, a center alignment loss is a… view at source ↗
Figure 3
Figure 3. Illustration of the token swapping, where node 1 is the target node. We first select node 3 and regard the tokens in its token sequences as the candidates. Then we select node 6 from the candidates to swap node 3, and construct the new token sequence. and restricts the diversity of the token sequences, further limiting the model performance. To effectively obtain diverse token sequences, SwapGT in￾troduces a novel o… view at source ↗
Figures from the paper (7 more)
Figure 4
Figure 4. Figure 4: Performances of SwapGT with or without the center alignment loss. the performance. Therefore, when the training data is suffi￾cient, GTs can achieve promising performance. And when the training data is sparse, GTs usually leg behind GNNs. Our proposed SwapGT addresses …
Figure 6
Figure 6. Figure 6: Analysis on the swapping times t. 1 2 3 4 5 6 7 8 90.00 92.00 94.00 96.00 98.00 1 2 3 4 5 6 7 8 80.00 85.00 90.00 95.00 100.00 1 2 3 4 5 6 7 8 80.00 85.00 90.00 95.00 100.00 1 2 3 4 5 6 7 8 50.00 60.00 70.00 80.00 90.00 Sparse Dense Photo Accuracy (%) s ACM Accuracy (%…
Figure 7
Figure 7. Figure 7: Analysis on the augmentation times s. eration can fully leverage the semantic relevance of nodes to generate informative token sequences. In this section, we evaluate the effectiveness of the proposed strategy by comparing it with two naive strategies. One is to enlarg…
Figure 8
Figure 8. Figure 8: Performances of SwapGT with or without the center alignment loss. B.2. Study of the token sequence generation The experimental results of SwapGT with different token sequence generation strategies on the rest datasets are shown in [PITH_FULL_IMAGE:figures/full_fig_p01…
Figure 9
Figure 9. Figure 9: We can find that the additional experimental results exhibit similar observations shown in the main text. This situation demonstrates the effectiveness of the token sequence generation with the proposed token swapping operation in enhancing the performance of tokenized…
Figure 10
Figure 10. Figure 10: Performances of SwapGT with varying t. 1 2 3 4 5 6 7 8 86.00 88.00 90.00 92.00 94.00 1 2 3 4 5 6 7 8 60.00 65.00 70.00 75.00 80.00 1 2 3 4 5 6 7 8 70.00 75.00 80.00 85.00 90.00 1 2 3 4 5 6 7 8 60.00 70.00 80.00 90.00 100.00 Sparse Dense Computer Accuracy (%) s Citesee…
Figure 11
Figure 11. Figure 11: Performances of SwapGT with varying s. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_11.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Rethinking Federated Graph Foundation Models: A Graph-Language Alignment-based Approach

    cs.LG 2026-01 conditional novelty 6.0 of 10

    FedGALA replaces vector-quantized federated graph foundation models with continuous graph-text contrastive alignment plus prompt tuning, claiming up to 14.37% gains over 22 baselines.

Reference graph

Works this paper leans on

43 extracted references · 39 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    V., and Galstyan, A

    Abu - El - Haija, S., Perozzi, B., Kapoor, A., Alipourfard, N., Lerman, K., Harutyunyan, H., Steeg, G. V., and Galstyan, A. Mixhop: Higher-order graph convolutional architectures via sparsified neighborhood mixing. In Proceedings of the International Conference on Machine Learning, 2019

    work page 2019

  3. [3]

    Beyond low-frequency information in graph convolutional networks

    Bo, D., Wang, X., Shi, C., and Shen, H. Beyond low-frequency information in graph convolutional networks. In Proceedings of the AAAI Conference on Artificial Intelligence , 2021

    work page 2021

  4. [4]

    Specformer: Spectral graph neural networks meet transformers

    Bo, D., Shi, C., Wang, L., and Liao, R. Specformer: Spectral graph neural networks meet transformers. In Proceedings of the International Conference on Learning Representations, 2023

    work page 2023

  5. [5]

    How attentive are graph attention networks? In Proceedings of the International Conference on Learning Representations, 2022

    Brody, S., Alon, U., and Yahav, E. How attentive are graph attention networks? In Proceedings of the International Conference on Learning Representations, 2022

    work page 2022

  6. [6]

    Measuring and relieving the over-smoothing problem for graph neural networks from the topological view

    Chen, D., Lin, Y., Li, W., Li, P., Zhou, J., and Sun, X. Measuring and relieving the over-smoothing problem for graph neural networks from the topological view. In Proceedings of the AAAI Conference on Artificial Intelligence , 2020 a

    work page 2020

  7. [7]

    Nagphormer: A tokenized graph transformer for node classification in large graphs

    Chen, J., Gao, K., Li, G., and He, K. Nagphormer: A tokenized graph transformer for node classification in large graphs. In Proceedings of the International Conference on Learning Representations, 2023 a

    work page 2023

  8. [8]

    E., and He, K

    Chen, J., Li, G., Hopcroft, J. E., and He, K. Signgt: Signed attention-based graph transformer for graph representation learning. CoRR, abs/2310.11025, 2023 b

    arXiv 2023

Show all 43 references
  1. [9]

    Ntformer: A composite node tokenized graph transformer for node classification

    Chen, J., Jiang, S., and He, K. Ntformer: A composite node tokenized graph transformer for node classification. CoRR, abs/2406.19249, 2024 a

    2024 arXiv

  2. [10]

    Neighborhood convolutional graph neural network

    Chen, J., Li, B., and He, K. Neighborhood convolutional graph neural network. Knowledge-Based Systems, pp.\ 111861, 2024 b

    2024

  3. [11]

    Pamt: A novel propagation-based approach via adaptive similarity mask for node classification

    Chen, J., Li, B., He, Q., and He, K. Pamt: A novel propagation-based approach via adaptive similarity mask for node classification. IEEE Transactions on Computational Social Systems, 2024 c

    2024

  4. [12]

    Nagphormer+: A tokenized graph transformer with neighborhood augmentation for node classification in large graphs

    Chen, J., Liu, C., Gao, K., Li, G., and He, K. Nagphormer+: A tokenized graph transformer with neighborhood augmentation for node classification in large graphs. IEEE Transactions on Big Data, 2024 d

    2024

  5. [13]

    Simple and deep graph convolutional networks

    Chen, M., Wei, Z., Huang, Z., Ding, B., and Li, Y. Simple and deep graph convolutional networks. In Proceedings of the International Conference on Machine Learning, 2020 b

    2020

  6. [14]

    Adaptive Universal Generalized PageRank Graph Neural Network

    Chien, E., Peng, J., Li, P., and Milenkovic, O. Adaptive Universal Generalized PageRank Graph Neural Network . In Proceedings of the International Conference on Learning Representations, 2021

    2021

  7. [15]

    Polynormer: Polynomial-expressive graph transformer in linear time

    Deng, C., Yue, Z., and Zhang, Z. Polynormer: Polynomial-expressive graph transformer in linear time. In Proceedings of the International Conference on Learning Representations, 2024

    2024

  8. [16]

    Vcr-graphormer: A mini-batch graph transformer via virtual connections

    Fu, D., Hua, Z., Xie, Y., Fang, J., Zhang, S., Sancak, K., Wu, H., Malevich, A., He, J., and Long, B. Vcr-graphormer: A mini-batch graph transformer via virtual connections. In Proceedings of the International Conference on Learning Representations, 2024

    2024

  9. [17]

    Block modeling-guided graph convolutional neural networks

    He, D., Liang, C., Liu, H., Wen, M., Jiao, P., and Feng, Z. Block modeling-guided graph convolutional neural networks. In Proceedings of the Thirty-Sixth AAAI Conference on Artificial Intelligence , 2022 a

    2022

  10. [18]

    Structural robust label propagation on homogeneous graphs

    He, Q., Chen, J., Xu, H., and He, K. Structural robust label propagation on homogeneous graphs. In Proceedings of the IEEE International Conference on Data Mining , 2022 b

    2022

  11. [19]

    Kipf, T. N. and Welling, M. Semi-supervised Classification with Graph Convolutional Networks . In Proceedings of the International Conference on Learning Representations, 2017

    2017

  12. [20]

    Predict then propagate: Graph neural networks meet personalized pagerank

    Klicpera, J., Bojchevski, A., and G \" u nnemann, S. Predict then propagate: Graph neural networks meet personalized pagerank. In Proceedings of the International Conference on Learning Representations, 2019

    2019

  13. [21]

    Finding global homophily in graph neural networks when meeting heterophily

    Li, X., Zhu, R., Cheng, Y., Shan, C., Luo, S., Li, D., and Qian, W. Finding global homophily in graph neural networks when meeting heterophily. In Proceedings of the International Conference on Machine Learning, 2022

    2022

  14. [22]

    Revisiting heterophily for graph neural networks

    Luan, S., Hua, C., Lu, Q., Zhu, J., Zhao, M., Zhang, S., Chang, X., and Precup, D. Revisiting heterophily for graph neural networks. In Proceedings of the Annual Conference on Neural Information Processing Systems, 2022

    2022

  15. [23]

    Polyformer: Scalable node-wise filters via polynomial graph transformer

    Ma, J., He, M., and Wei, Z. Polyformer: Scalable node-wise filters via polynomial graph transformer. In Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2024

    2024

  16. [24]

    Rethinking structural encodings: Adaptive graph transformer for node classification task

    Ma, X., Chen, Q., Wu, Y., Song, G., Wang, L., and Zheng, B. Rethinking structural encodings: Adaptive graph transformer for node classification task. In Proceedings of the ACM Web Conference , 2023

    2023

  17. [25]

    Co-embedding attributed networks

    Meng, Z., Liang, S., Bao, H., and Zhang, X. Co-embedding attributed networks. In Proceedings of the Twelfth ACM International Conference on Web Search and Data Mining , 2019

    2019

  18. [26]

    C., Lei, Y., and Yang, B

    Pei, H., Wei, B., Chang, K. C., Lei, Y., and Yang, B. Geom-gcn: Geometric graph convolutional networks. In Proceedings of the International Conference on Learning Representations, 2020

    2020

  19. [27]

    Platonov, O., Kuznedelev, D., Diskin, M., Babenko, A., and Prokhorenkova, L. A critical look at the evaluation of gnns under heterophily: Are we really making progress? In Proceedings of the Eleventh International Conference on Learning Representations, 2023

    2023

  20. [28]

    P., Luu, A

    Ramp \'a s ek, L., Galkin, M., Dwivedi, V. P., Luu, A. T., Wolf, G., and Beaini, D. Recipe for a general, powerful, scalable graph transformer. In Proceedings of the Annual Conference on Neural Information Processing Systems, 2022

    2022

  21. [29]

    N., Kaiser, ., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention Is All You Need . In Proceedings of the Annual Conference on Neural Information Processing Systems, 2017

    2017

  22. [30]

    Graph Attention Networks

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Lio, P., and Bengio, Y. Graph Attention Networks . In Proceedings of the International Conference on Learning Representations, 2018

    2018

  23. [31]

    Wang, X., Ji, H., Shi, C., Wang, B., Ye, Y., Cui, P., and Yu, P. S. Heterogeneous graph attention network. In Proceedings of the World Wide Web Conference, 2019

    2019

  24. [32]

    AM-GCN: adaptive multi-channel graph convolutional networks

    Wang, X., Zhu, M., Bo, D., Cui, P., Shi, C., and Pei, J. AM-GCN: adaptive multi-channel graph convolutional networks. In Proceedings of the ACM SIGKDD Conference on Knowledge Discovery and Data Mining , 2020

    2020

  25. [33]

    Simplifying Graph Convolutional Networks

    Wu, F., Souza, A., Zhang, T., Fifty, C., Yu, T., and Weinberger, K. Simplifying Graph Convolutional Networks . In Proceedings of the International Conference on Machine Learning, 2019

    2019

  26. [34]

    Nodeformer: A scalable graph structure learning transformer for node classification

    Wu, Q., Zhao, W., Li, Z., Wipf, D., and Yan, J. Nodeformer: A scalable graph structure learning transformer for node classification. In Proceedings of the Annual Conference on Neural Information Processing Systems, 2022

    2022

  27. [35]

    Simplifying and empowering transformers for large-graph representations

    Wu, Q., Zhao, W., Yang, C., Zhang, H., Nie, F., Jiang, H., Bian, Y., and Yan, J. Simplifying and empowering transformers for large-graph representations. In Proceedings of the Annual Conference on Neural Information Processing Systems, 2023

    2023

  28. [36]

    Less is more: on the over-globalizing problem in graph transformers

    Xing, Y., Wang, X., Li, Y., Huang, H., and Shi, C. Less is more: on the over-globalizing problem in graph transformers. In Proceedings of the International Conference on Machine Learning, 2024

    2024

  29. [37]

    Representation learning on graphs with jumping knowledge networks

    Xu, K., Li, C., Tian, Y., Sonobe, T., Kawarabayashi, K.-i., and Jegelka, S. Representation learning on graphs with jumping knowledge networks. In Proceedings of the International conference on machine learning, 2018

    2018

  30. [38]

    FPGNN: fair path graph neural network for mitigating discrimination

    Zhang, G., Cheng, D., and Zhang, S. FPGNN: fair path graph neural network for mitigating discrimination. World Wide Web, 26 0 (5): 0 3119--3136, 2023

    2023

  31. [39]

    Learning fair representations via rebalancing graph structure

    Zhang, G., Cheng, D., Yuan, G., and Zhang, S. Learning fair representations via rebalancing graph structure. Inf. Process. Manag., 61 0 (1): 0 103570, 2024

    2024

  32. [40]

    Disentangled contrastive learning for fair graph representations

    Zhang, G., Yuan, G., Cheng, D., Liu, L., Li, J., and Zhang, S. Disentangled contrastive learning for fair graph representations. Neural Networks, 181: 0 106781, 2025

    2025

  33. [41]

    Hierarchical Graph Transformer with Adaptive Node Sampling

    Zhang, Z., Liu, Q., Hu, Q., and Lee, C. Hierarchical Graph Transformer with Adaptive Node Sampling . In Proceedings of the Annual Conference on Neural Information Processing Systems, 2022

    2022

  34. [42]

    Gophormer: Ego-Graph Transformer for Node Classification

    Zhao, J., Li, C., Wen, Q., Wang, Y., Liu, Y., Sun, H., Xie, X., and Ye, Y. Gophormer: Ego-Graph Transformer for Node Classification . arXiv preprint arXiv:2110.13094, 2021

    2021 arXiv

  35. [43]

    Beyond homophily in graph neural networks: Current limitations and effective designs

    Zhu, J., Yan, Y., Zhao, L., Heimann, M., Akoglu, L., and Koutra, D. Beyond homophily in graph neural networks: Current limitations and effective designs. In Proceedings of the Annual Conference on Neural Information Processing Systems, 2020

    2020

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.