Pith. sign in

REVIEW 4 major objections 5 minor 47 references

Aggregation Buffer: Revisiting DropEdge with a New Parameter Block

T0 review · 4 major / 5 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 A small add-on block lets DropEdge finally improve GNN accuracy and robustness.

desk verdict A practical AGGB module with broad empirical support carries a theory section that overclaims: the proofs do not establish that DropEdge fails due to a fundamental GNN limitation. read the letter →

arxiv 2505.20840 v2 pith:3W7IZLMD submitted 2025-05-27 cs.LG

Dooho Lee , Myeong Kong , Sagad Hamid , Cheonwoo Lee , Jaemin Yoo This is my paper
classification cs.LG
keywords graphneuralnetworksDropEdgerobustnessdegreebiasstructuraldisparityaggregationbuffernodeclassificationdataaugmentation
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

DropEdge, the standard trick of randomly deleting edges during graph neural network training, barely improves test accuracy — and this paper argues that the fault is not DropEdge itself but the GNN architecture it is applied to. The paper proves that GNN layers whose aggregation depends on the adjacency matrix cannot have an input-independent bound on how much their outputs change when edges are dropped, so robustness under structural perturbation cannot be optimized by DropEdge training alone. To repair this, the paper adds a cheap degree-normalized parameter block, called Aggregation Buffer (AGGB), that is trained after the base GNN is frozen, using a robustness-controlled loss with DropEdge. Across 12 node-classification benchmarks and several GNN architectures, the augmented models consistently beat both the base GNNs and DropEdge alone, with the largest gains on low-degree and heterophilous nodes.

What carries the argument

Three pieces carry the argument. First, the root subgraph view: each node's prediction depends on a rooted subgraph, and DropEdge produces reduced rooted subgraphs, shifting the training objective into bias plus robustness terms (the KL divergence between the model's outputs with and without dropped edges). Second, discrepancy bounds: MLPs have a clean input-independent composition bound ($C = L_\sigma \prod \|W_i\|_2$), while GCNs cannot have such a constant bound; instead, a GCN layer obeys $\|H^{(l)}_1 - H^{(l)}_2\|_2 \le C_1\|H^{(l-1)}_1 - H^{(l-1)}_2\|_2 + C_2$ with $C_2 = C_1|V|\|\hat A_1 - \hat A_2\|_2$, an additive term that arises purely from the adjacency change. Third, the proposed AGGB block $g_B(H^{(0:l-1)}, A) = (D+I)^{-1}H^{(0:l-1)}W^{(l)}$, which is edge-aware and stable, so its output magnitude increases when edges are dropped, compensating for the perturbation.

What would settle it

Track the bias and robustness losses (and test accuracy) during joint training of a GNN with DropEdge: if some training schedule or loss design achieves both low robustness loss and low bias loss without AGGB, the claim that the architecture fundamentally prevents optimizing the robustness term would be falsified.

Watch

Extended reading notes

Core claim

The paper claims that the failure of DropEdge is not rooted in its algorithm but in the inductive bias of GNNs: because the aggregation operation depends on the specific adjacency matrix, perturbing edges injects an irreducible additive discrepancy term that no constant input-independent bound can eliminate. The overarching claim is that a deliberately placed, degree-normalized additive block — Aggregation Buffer — trained on a frozen pretrained GNN with a robustness-controlled loss repairs this architectural limitation, yielding consistent gains in accuracy, degree bias, and structural disparity across architectures and datasets.

Load-bearing premise

The paper assumes that the additive discrepancy term in the GNN bound is what actually prevents gradient descent from optimizing the robustness objective, rather than merely describing a bound that is loose in practice.

Editorial extensions

If this is right

  • DropEdge becomes a reliable post-hoc robustness regularizer when combined with AGGB, instead of a data augmentation that merely trades robustness for bias.
  • AGGB attaches to already-deployed GNNs: training only the buffer parameters on a frozen backbone means the original model can be recovered by detaching the block.
  • Edge-robustness is positioned as a unifying mechanism behind two known GNN failure modes — degree bias and structural disparity — with AGGB improving low-degree and heterophilous nodes in most tested datasets.
  • Deeper GNNs benefit more from AGGB, suggesting that repeated aggregation accumulates the additive discrepancy term, and that edge-robustness may partly explain oversmoothing degradation.
  • AGGB generalizes across architectures (SAGE, GAT, SGC, GIN), consistent with the claim that the limitation is shared across GNN inductive biases.

Reading between the lines

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

  • The theory suggests that any GNN aggregation layer whose normalization depends on the adjacency matrix will carry a similar additive discrepancy, so AGGB should help other message-passing variants (e.g., graph transformers with attention on edges) even though the paper only proves SAGE, GIN, and GAT bounds.
  • Because AGGB's robustness term uses unlabeled nodes, the method should be especially effective in semi-supervised regimes with few labels — a testable prediction the paper does not emphasize.
  • One could test the claim more sharply by measuring whether a node whose rooted subgraph undergoes a large additive perturbation $\|\hat A_1 - \hat A_2\|_2$ also shows the largest accuracy gains from AGGB, linking the bound to per-node benefit.
  • If the discrepancy bound is the true cause, then pre-training the backbone with harder edge perturbations or a curriculum over drop ratios should improve AGGB further; if it does not, the bound may be only loosely tied to optimization dynamics.
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

4 major / 5 minor

Summary. The paper revisits DropEdge for GNNs and argues that its limited performance gains are not due to the DropEdge algorithm itself but to an architectural inductive bias of GNNs. The authors formalize a bias-robustness decomposition of the DropEdge objective, prove discrepancy bounds showing that GNNs lack an input-independent Lipschitz constant when adjacency matrices differ, and attribute DropEdge's failure to the additive adjacency-dependent term in these bounds. They then propose Aggregation Buffer (AGGB), a degree-normalized additive block inserted into a frozen pretrained GNN and trained with a robustness-controlled loss under DropEdge. Experiments across 12 node-classification datasets, several GNN architectures, and ablations report consistent accuracy gains, improved tail-node performance, and better robustness under random edge removal.

Significance. If the central claim were established, the paper would provide a useful diagnosis of DropEdge and a modular, architecture-agnostic remedy with practical value: AGGB can be attached to an already-trained GNN, preserves the pretrained model, and improves accuracy and degree-bias metrics across many benchmarks. The empirical study is extensive and carefully ablated, including comparisons to random-dropping baselines, degree-bias methods, alternative layer designs, alternative losses, deeper architectures, and larger datasets, and the code is promised to be released. The main weakness is that the theoretical explanation, which is the paper's headline contribution, is not actually derived: the stated theorems do not imply that the robustness objective is unoptimizable, and the proposed block does not escape the same discrepancy phenomenon. The empirical method may still be valid, but the paper currently overclaims the theoretical foundations.

major comments (4)
  1. [Section 3.4, Theorems 3.8 and 3.9] The central claim that the additive term C2 in Theorem 3.9 'hinders the optimization of the robustness term' is not proved. Theorem 3.8 only establishes that no input-independent Lipschitz constant exists when adjacency matrices differ, and Theorem 3.9 provides an upper bound; neither statement implies that gradient descent cannot find parameters that nearly satisfy the robustness objective on the observed distribution of reduced rooted subgraphs. In fact, Figure 2 shows that DropEdge does decrease the robustness term, so C2 does not empirically prevent robustness optimization. The paper needs either an explicit optimization-theoretic obstruction or a substantially weakened claim that the explanation is a sufficient, not necessary, qualitative account.
  2. [Section 3.3 and Figure 2] The observed bias increase under DropEdge is attributed to the additive discrepancy term, but the same pattern would be expected from ordinary regularization or from training on a perturbed objective: the model is pulled away from the minimizer of the original loss. The paper does not provide a control experiment or an intermediate-quantity analysis to show that the bias rise is specifically caused by the adjacency-dependent C2 term rather than by a generic bias-variance trade-off. A control using the same two-term loss with an input-independent augmentation, or tracking intermediate representation distances during training, would directly test the proposed mechanism.
  3. [Section 4.2, Theorem 4.1, and Section 4.3] AGGB does not escape the theoretical limitation claimed in Theorems 3.8 and 3.9: the buffer gB depends on the adjacency matrix through (D+I)^{-1}, so the augmented network GNNB still has no input-independent discrepancy constant. Conditions C1 and C2 are qualitative, and Theorem 4.1 only proves a Frobenius-norm inequality between the buffer outputs on A and on the perturbed adjacency matrix; it does not establish that the robustness term in LRC is optimizable or that AGGB 'repairs' the architectural limitation. The theoretical justification of AGGB is therefore not supported by the analysis as written, even though the empirical results may still justify the method.
  4. [Appendix F and Table 10] The edge-removal robustness evaluation measures, at test time, essentially the same quantity that the proposed LRC loss is trained to optimize, so part of the reported robustness gain is by construction. The overall-accuracy and degree-bias results are external benchmarks and are not affected by this circularity, but the paper's broader claim of 'edge-robustness' as demonstrated by Table 10 would be stronger if it also evaluated perturbations not used during training, such as adversarial edge deletions or a different augmentation distribution.
minor comments (5)
  1. [Section 3.2 and Figure 2] The bias term DKL(P(yi|Gi)||Q(yi|Gi)) is not directly computable from finite labels, yet Figure 2 reports it on test data; the paper should specify the estimator used and its variance.
  2. [Table 1 and main text, Section 6.1] The caption says GCNB achieves at least second-best in 31 of 36 settings, while the text says 9 first- and 3 second-best in 12 overall rows; the counting criteria for the 36 settings should be stated explicitly so the two claims are consistent.
  3. [Introduction, page 2] There is a typo in 'which can be integrated to to any trained GNN'.
  4. [Appendix C.4] The proof of Lemma 3.7 assumes a normalization condition on H, namely ||H*||2 <= |V|, which is not stated in the main text or in the lemma; this assumption should be moved to the main statement.
  5. [Section 4.3] The switch from E_P in Equation (3) to DKL(Q||Q) in Equation (4) is justified only heuristically in Appendix E; the paper should either provide a rigorous condition under which the approximation is valid or clearly label it as an empirical modeling choice.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the theoretical claims are derived from stated Lipschitz properties, and the AGGB design is an explicit construction rather than a renamed input, though the robustness evaluation closely matches the training objective.

full rationale

The paper's derivation chain is self-contained. The core theoretical results (Theorems 3.8 and 3.9) are derived from the Lipschitz continuity of activations and aggregation operations, as shown in Appendix C; the non-existence of an input-independent discrepancy bound for different adjacency matrices is a standard linear-algebra argument and does not assume the conclusion that GNNs cannot optimize the robustness term. The proposed AGGB block is explicitly constructed to satisfy the paper's own conditions C1 and C2, and Theorem 4.1 verifies this by a direct norm inequality; this is a design proof, not a circular reduction. The training objective LRC in Equation (5) explicitly includes a robustness term DKL(QB(yi|Gi) || QB(yi|~Gi)), so improved accuracy under edge removal at test time is a predictable consequence of the chosen loss, but this is a standard train/eval loop rather than a claim that the result was predicted from first principles. The paper's broader accuracy, degree-bias, and structural-disparity results are evaluated on held-out nodes against external baselines, giving them independent empirical content. There are no load-bearing self-citations: the references are to prior external work, and no uniqueness theorem or prior result by the same authors is invoked to force the choice of AGGB. The main weakness identified by the skeptic, namely that Theorem 3.9's upper bound does not logically imply that gradient descent cannot optimize the robustness term, is a soundness or sufficiency gap in the explanation, not a case where the paper's conclusion is equivalent by definition to its inputs. Accordingly, no specific circular step can be quoted and exhibited under the rules, and the appropriate score is 0.

Assumptions & free parameters 5 free parameters · 6 assumptions · 1 invented entities

The method's claims rest mostly on the standard supervised classification setup, plus the Q approximately P approximation used to turn DropEdge's objective into a computable bias and robustness loss, plus the author-defined C2 stability condition. The learned AGGB weights and per-dataset hyperparameters are the main free parameters.

free parameters (5)
  • AGGB weight matrices W_B per layer = learned during second-stage training, not reported per entry
    These matrices define the AGGB correction term and are fitted on each dataset.
  • Lambda balancing bias and robustness loss = per dataset, Table 9, values 0.1, 0.5, or 1.0
    Selected by validation accuracy in a grid search over [1, 0.5, 0.1].
  • DropEdge ratio for AGGB training = per dataset, Table 9, values 0.2 to 0.7
    Chosen by validation accuracy in a grid search over [0.2, 0.5, 0.7, 1.0].
  • Dropout ratio for AGGB training = per dataset, Table 9, values 0.0 to 0.7
    Chosen by validation accuracy in a grid search over [0, 0.2, 0.5, 0.7].
  • Base GCN hyperparameters = per dataset, Table 8, including hidden dimension, learning rate, weight decay, dropout, aggregation scheme
    The pretrained backbone is selected by grid search, and AGGB is attached to it.
assumptions (6)
  • domain assumption The true label distribution P can be approximated by the pretrained model distribution Q in Equation (4) and in the bias term of LRC.
    The paper discusses this in Appendix E but does not verify the approximation on test nodes.
  • domain assumption Bias and robustness terms in the shifted DropEdge objective are separable and can be optimized independently in two stages.
    This underpins the two-stage training scheme; no formal optimization guarantee is provided.
  • standard math Activation functions are non-constant, continuous, and element-wise for Lemma C.1.
    ReLU, Sigmoid, and GELU satisfy this, and the lemma requires it to construct a separating input.
  • ad hoc to paper The Frobenius norm condition C2 is a sufficient proxy for preserving pretrained knowledge.
    C2 is an author-defined design condition; the paper does not prove that satisfying it is necessary or sufficient for accuracy preservation.
  • domain assumption Test-time random edge removal is a meaningful proxy for the structural inconsistencies the method targets.
    The robustness evaluation in Appendix F assumes that random edge deletion reflects degree bias and structural disparity.
  • standard math Standard undirected graph and normalized adjacency assumptions from GCN hold.
    The analysis uses row-normalized and symmetric-normalized adjacency matrices in the standard way.
invented entities (1)
  • Aggregation Buffer (AGGB) parameter block independent evidence
    purpose: An additive, degree-normalized layer that maps all previous representations to a correction term, trained after freezing the backbone GNN to improve edge robustness.
    The block is a computational component with released code and benchmark results, so its behavior can be independently tested outside this paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Aggregation Buffer: Revisiting DropEdge with a New Parameter Block." pith.science (2026). https://pith.science/paper/3W7IZLMD

@misc{pith2026250520840,
  author       = {Pith},
  title        = {Pith review of: Aggregation Buffer: Revisiting DropEdge with a New Parameter Block},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3W7IZLMD}},
  note         = {Machine review of arXiv:2505.20840}
}
read the original abstract

We revisit DropEdge, a data augmentation technique for GNNs which randomly removes edges to expose diverse graph structures during training. While being a promising approach to effectively reduce overfitting on specific connections in the graph, we observe that its potential performance gain in supervised learning tasks is significantly limited. To understand why, we provide a theoretical analysis showing that the limited performance of DropEdge comes from the fundamental limitation that exists in many GNN architectures. Based on this analysis, we propose Aggregation Buffer, a parameter block specifically designed to improve the robustness of GNNs by addressing the limitation of DropEdge. Our method is compatible with any GNN model, and shows consistent performance improvements on multiple datasets. Moreover, our method effectively addresses well-known problems such as degree bias or structural disparity as a unifying solution. Code and datasets are available at https://github.com/dooho00/agg-buffer.

Figures

Figures reproduced from arXiv: 2505.20840 by the authors.

Figure 1
Figure 1. DropEdge generates various reduced rooted subgraphs for center nodes (*) by randomly removing edges. In node-level tasks, a GNN can be considered as taking the k-hop subgraph of each node as its input. For each node i, the edge removal operation in DropEdge can be interpreted as transforming the rooted subgraph Gi , centered on node i, into a reduced rooted subgraph, denoted as G˜ i [PITH_FULL_IMAGE:figures/full_f… view at source ↗
Figure 2
Figure 2. Accuracy and loss terms on test data during the training of a GCN at PubMed. We illustrate the average of 10 independent runs, with shaded regions representing the minimum and maximum values. While DropEdge decreases the robustness term compared to standard GNNs, it leads to increasing the bias term, eventually resulting in similar test accuracy to standard GNNs. predictions for Gi and G˜ i . However, as the expecta… view at source ↗
Figure 3
Figure 3. AGGB is specifically designed to refine the output of the AGG operation, mitigating discrepancies caused by variations in the graph structure introduced by DropEdge. 4.1. Aggregation Buffer: A New Parameter Block Unlike the standard training strategy, where an augmenta￾tion function is used during training, we propose a two-step approach; given a GNN trained without DropEdge, we in￾tegrate AGGB into each GNN layer a… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Changes in two different approximations of the robustness loss, EP [log Q(yi|Gi) − log Q(yi|G˜i)], during training of the base GCN (top row) and AGGB (bottom row). Each curve represents the average over 10 independent runs, with shaded areas indicating the minimum and …

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

47 extracted references · 26 canonical work pages

  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]

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

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

    2022

  3. [3]

    Simple and deep graph convolutional networks

    Chen, M., Wei, Z., Huang, Z., Ding, B., and Li, Y. Simple and deep graph convolutional networks. In International conference on machine learning, pp.\ 1725--1735. PMLR, 2020

    work page 2020

  4. [4]

    Improved regularization of convolutional neural networks with cutout

    DeVries, T. Improved regularization of convolutional neural networks with cutout. arXiv preprint arXiv:1708.04552, 2017

    arXiv 2017

  5. [5]

    P., Joshi, C

    Dwivedi, V. P., Joshi, C. K., Luu, A. T., Laurent, T., Bengio, Y., and Bresson, X. Benchmarking graph neural networks. Journal of Machine Learning Research, 24 0 (43): 0 1--48, 2023

    2023

  6. [6]

    Dropmessage: Unifying random dropping for graph neural networks

    Fang, T., Xiao, Z., Wang, C., Xu, J., Yang, X., and Yang, Y. Dropmessage: Unifying random dropping for graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp.\ 4267--4275, 2023

    work page 2023

  7. [7]

    Graph random neural networks for semi-supervised learning on graphs

    Feng, W., Zhang, J., Dong, Y., Han, Y., Luan, H., Xu, Q., Yang, Q., Kharlamov, E., and Tang, J. Graph random neural networks for semi-supervised learning on graphs. Advances in neural information processing systems, 33: 0 22092--22103, 2020

    work page 2020

  8. [8]

    Combining neural networks with personalized pagerank for classification on graphs

    Gasteiger, J., Bojchevski, A., and Günnemann, S. Combining neural networks with personalized pagerank for classification on graphs. In International Conference on Learning Representations, 2019

    work page 2019

Show all 47 references
  1. [9]

    S., Riley, P

    Gilmer, J., Schoenholz, S. S., Riley, P. F., Vinyals, O., and Dahl, G. E. Neural message passing for quantum chemistry. In International conference on machine learning, pp.\ 1263--1272. PMLR, 2017

    2017

  2. [10]

    Inductive representation learning on large graphs

    Hamilton, W., Ying, Z., and Leskovec, J. Inductive representation learning on large graphs. Advances in neural information processing systems, 30, 2017

    2017

  3. [11]

    Deep residual learning for image recognition

    He, K., Zhang, X., Ren, S., and Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE conference on computer vision and pattern recognition, pp.\ 770--778, 2016

    2016

  4. [12]

    Distilling the knowledge in a neural network

    Hinton, G. Distilling the knowledge in a neural network. arXiv preprint arXiv:1503.02531, 2015

    2015 arXiv

  5. [13]

    Y., Zhou, T., Wu, Y., Song, X., Song, X., and Zhou, D

    Hou, L., Pang, R. Y., Zhou, T., Wu, Y., Song, X., Song, X., and Zhou, D. Token dropping for efficient BERT pretraining. In Muresan, S., Nakov, P., and Villavicencio, A. (eds.), Proceedings of the 60th Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

    2022 doi

  6. [14]

    Open graph benchmark: Datasets for machine learning on graphs

    Hu, W., Fey, M., Zitnik, M., Dong, Y., Ren, H., Liu, B., Catasta, M., and Leskovec, J. Open graph benchmark: Datasets for machine learning on graphs. Advances in neural information processing systems, 33: 0 22118--22133, 2020 a

    2020

  7. [15]

    Strategies for pre-training graph neural networks

    Hu, W., Liu, B., Gomes, J., Zitnik, M., Liang, P., Pande, V., and Leskovec, J. Strategies for pre-training graph neural networks. In International Conference on Learning Representations, 2020 b

    2020

  8. [16]

    W., Subbian, K., and Leskovec, J

    Hu, W., Cao, K., Huang, K., Huang, E. W., Subbian, K., and Leskovec, J. Tuneup: A training strategy for improving generalization of graph neural networks, 2023. URL https://openreview.net/forum?id=8xuFD1yCoH

    2023

  9. [17]

    Graphpatcher: mitigating degree bias for graph neural networks via test-time augmentation

    Ju, M., Zhao, T., Yu, W., Shah, N., and Ye, Y. Graphpatcher: mitigating degree bias for graph neural networks via test-time augmentation. Advances in Neural Information Processing Systems, 36, 2024

    2024

  10. [18]

    Kingma, D. P. and Ba, J. Adam: A method for stochastic optimization. In Bengio, Y. and LeCun, Y. (eds.), 3rd International Conference on Learning Representations, ICLR 2015, San Diego, CA, USA, May 7-9, 2015, Conference Track Proceedings , 2015

    2015

  11. [19]

    Kipf, T. N. and Welling, M. Semi-supervised classification with graph convolutional networks. In 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings . OpenReview.net, 2017

    2017

  12. [20]

    Crafting papers on machine learning

    Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML 2000), pp.\ 1207--1216, Stanford, CA, 2000. Morgan Kaufmann

    2000

  13. [21]

    Lee, D.-H. et al. Pseudo-label: The simple and efficient semi-supervised learning method for deep neural networks. In Workshop on challenges in representation learning, ICML, volume 3, pp.\ 896. Atlanta, 2013

    2013

  14. [22]

    What's behind the mask: Understanding masked graph modeling for graph autoencoders

    Li, J., Wu, R., Sun, W., Chen, L., Tian, S., Zhu, L., Meng, C., Zheng, Z., and Wang, W. What's behind the mask: Understanding masked graph modeling for graph autoencoders. In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pp.\ 1268--1279, 2023

    2023

  15. [23]

    Tail-gnn: Tail-node graph neural networks

    Liu, Z., Nguyen, T.-K., and Fang, Y. Tail-gnn: Tail-node graph neural networks. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining, pp.\ 1109--1119, 2021

    2021

  16. [24]

    On generalized degree fairness in graph neural networks

    Liu, Z., Nguyen, T.-K., and Fang, Y. On generalized degree fairness in graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 37, pp.\ 4525--4533, 2023

    2023

  17. [25]

    Classic GNN s are strong baselines: Reassessing GNN s for node classification

    Luo, Y., Shi, L., and Wu, X.-M. Classic GNN s are strong baselines: Reassessing GNN s for node classification. In The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track, 2024

    2024

  18. [26]

    Demystifying structural disparity in graph neural networks: Can one size fit all? Advances in neural information processing systems, 36, 2024

    Mao, H., Chen, Z., Jin, W., Han, H., Ma, Y., Zhao, T., Shah, N., and Tang, J. Demystifying structural disparity in graph neural networks: Can one size fit all? Advances in neural information processing systems, 36, 2024

    2024

  19. [27]

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

    Pei, H., Wei, B., Chang, K. C.-C., Lei, Y., and Yang, B. Geom-gcn: Geometric graph convolutional networks. arXiv preprint arXiv:2002.05287, 2020

    2002 arXiv

  20. [28]

    A critical look at the evaluation of GNN s under heterophily: Are we really making progress? In The Eleventh International Conference on Learning Representations, 2023

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

    2023

  21. [29]

    Dropedge: Towards deep graph convolutional networks on node classification

    Rong, Y., Huang, W., Xu, T., and Huang, J. Dropedge: Towards deep graph convolutional networks on node classification. In International Conference on Learning Representations, 2020

    2020

  22. [30]

    Pitfalls of graph neural network evaluation

    Shchur, O., Mumme, M., Bojchevski, A., and G \"u nnemann, S. Pitfalls of graph neural network evaluation. arXiv preprint arXiv:1811.05868, 2018

    2018 arXiv

  23. [31]

    Graph-based semi-supervised learning: A comprehensive review

    Song, Z., Yang, X., Xu, Z., and King, I. Graph-based semi-supervised learning: A comprehensive review. IEEE Transactions on Neural Networks and Learning Systems, 34 0 (11): 0 8174--8194, 2022

    2022

  24. [32]

    Dropout: a simple way to prevent neural networks from overfitting

    Srivastava, N., Hinton, G., Krizhevsky, A., Sutskever, I., and Salakhutdinov, R. Dropout: a simple way to prevent neural networks from overfitting. The journal of machine learning research, 15 0 (1): 0 1929--1958, 2014

    1929

  25. [33]

    Theoretical and empirical insights into the origins of degree bias in graph neural networks

    Subramonian, A., Kang, J., and Sun, Y. Theoretical and empirical insights into the origins of degree bias in graph neural networks. In The Thirty-eighth Annual Conference on Neural Information Processing Systems, 2024

    2024

  26. [34]

    Investigating and mitigating degree-related biases in graph convoltuional networks

    Tang, X., Yao, H., Sun, Y., Wang, Y., Tang, J., Aggarwal, C., Mitra, P., and Wang, S. Investigating and mitigating degree-related biases in graph convoltuional networks. In Proceedings of the 29th ACM International Conference on Information & Knowledge Management, pp.\ 1435--1...

    2020

  27. [35]

    Graph attention networks

    Veli c kovi \'c , P., Cucurull, G., Casanova, A., Romero, A., Li \`o , P., and Bengio, Y. Graph attention networks. In International Conference on Learning Representations, 2018

    2018

  28. [36]

    Understanding heterophily for graph neural networks

    Wang, J., Guo, Y., Yang, L., and Wang, Y. Understanding heterophily for graph neural networks. In Salakhutdinov, R., Kolter, Z., Heller, K., Weller, A., Oliver, N., Scarlett, J., and Berkenkamp, F. (eds.), Proceedings of the 41st International Conference on Machine Learning, v...

    2024

  29. [37]

    Deep graph library: A graph-centric, highly-performant package for graph neural networks

    Wang, M., Zheng, D., Ye, Z., Gan, Q., Li, M., Song, X., Zhou, J., Ma, C., Yu, L., Gai, Y., et al. Deep graph library: A graph-centric, highly-performant package for graph neural networks. arXiv preprint arXiv:1909.01315, 2019

    1909 arXiv

  30. [38]

    Simplifying graph convolutional networks

    Wu, F., Souza, A., Zhang, T., Fifty, C., Yu, T., and Weinberger, K. Simplifying graph convolutional networks. In International conference on machine learning, pp.\ 6861--6871. PMLR, 2019

    2019

  31. [39]

    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 International conference on machine learning, pp.\ 5453--5462. PMLR, 2018

    2018

  32. [40]

    How powerful are graph neural networks? In International Conference on Learning Representations, 2019

    Xu, K., Hu, W., Leskovec, J., and Jegelka, S. How powerful are graph neural networks? In International Conference on Learning Representations, 2019

    2019

  33. [41]

    L., and Leskovec, J

    Ying, R., He, R., Chen, K., Eksombatchai, P., Hamilton, W. L., and Leskovec, J. Graph convolutional neural networks for web-scale recommender systems. In Proceedings of the 24th ACM SIGKDD international conference on knowledge discovery & data mining, pp.\ 974--983, 2018

    2018

  34. [42]

    Graph contrastive learning with augmentations

    You, Y., Chen, T., Sui, Y., Chen, T., Wang, Z., and Shen, Y. Graph contrastive learning with augmentations. Advances in neural information processing systems, 33: 0 5812--5823, 2020

    2020

  35. [43]

    GraphSAINT : Graph sampling based inductive learning method

    Zeng, H., Zhou, H., Srivastava, A., Kannan, R., and Prasanna, V. GraphSAINT : Graph sampling based inductive learning method. In International Conference on Learning Representations, 2020

    2020

  36. [44]

    Zhang, H., Wu, Q., Yan, J., Wipf, D., and Yu, P. S. From canonical correlation analysis to self-supervised graph neural networks. Advances in Neural Information Processing Systems, 34: 0 76--89, 2021

    2021

  37. [45]

    Graph-less neural networks: Teaching old mlps new tricks via distillation

    Zhang, S., Liu, Y., Sun, Y., and Shah, N. Graph-less neural networks: Teaching old mlps new tricks via distillation. In International Conference on Learning Representations, 2022

    2022

  38. [46]

    W., Rao, N., Katariya, S., Wang, Z., and Subbian, K

    Zheng, W., Huang, E. W., Rao, N., Katariya, S., Wang, Z., and Subbian, K. Cold brew: Distilling graph node representations with incomplete or missing neighborhoods. arXiv preprint arXiv:2111.04840, 2021

    2021 arXiv

  39. [47]

    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. Advances in neural information processing systems, 33: 0 7793--7804, 2020

    2020

Pith tools

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