Pith. sign in

REVIEW 4 major objections 4 minor 53 references

ADMP-GNN: Adaptive Depth Message Passing GNN

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read The paper claims that the right number of message-passing layers is per-node, and that a centrality-based exit policy lets a single multi-exit GNN beat fixed-depth baselines.

desk verdict The sequential training trick is solid, but the headline claim of adaptive-depth gains does not survive a close reading of the tables. read the letter →

arxiv 2509.01170 v1 pith:XOBX443H submitted 2025-09-01 cs.LG stat.ML

Yassine Abbahaddou , Fragkiskos D. Malliaros , Johannes F. Lutzeyer , Michalis Vazirgiannis This is my paper
classification cs.LGstat.ML MSC 68T0768R10
keywords GraphNeuralNetworksNodeclassificationAdaptivedepthMessagepassingEarlyexitSequentialtrainingCentralityDynamic
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

Graph neural networks typically run every node through the same number of message-passing layers. This paper argues that different nodes are better classified at different depths: a node in a dense neighborhood can be predicted early, while a node in a sparse periphery needs more propagation. To exploit that, ADMP-GNN attaches a small classifier to every layer of any message-passing GNN and trains the layers one at a time, freezing earlier ones to avoid gradient conflict. A heuristic then groups nodes by centrality and assigns each group the exit layer that works best on validation nodes. The paper reports consistent accuracy gains over fixed-depth GCN and GIN baselines, with an oracle upper bound several points higher, indicating room to improve the policy.

What carries the argument

Two architectural additions to a message-passing GNN: an exit update that maps the aggregated message at layer l to a softmax prediction, and a continuation update that produces the hidden state for the next layer. Training is sequential: layer l's weights are optimized for the exit loss at l, then frozen, so each layer is learned to be useful both for its own exit and as input to the next. The policy layer is a centrality-bucket rule: nodes are sorted by degree, k-core, PageRank, or walk count, put into C equal-size buckets, and each bucket receives the exit layer with best validation accuracy.

What would settle it

On any benchmark dataset, compute the optimal exit layer for each test node, then check whether the distribution of optimal exits within centrality buckets differs from the global distribution; if it does not, centrality carries no signal. Simpler and more direct: shuffle the bucket assignments and see whether the policy accuracy stays the same, which would show the selection rule is not actually using structural information.

Watch

Extended reading notes

Core claim

The central claim is that the right number of message-passing steps is a per-node quantity, not a global hyperparameter. ADMP-GNN realizes this by making each layer produce a prediction through an exit update, while a continuation update keeps propagating representations for nodes that go deeper. The sequential training scheme trains each layer's parameters only against its own exit loss and then freezes them, which the paper shows avoids the gradient conflicts that make naive multi-exit training unstable and closely reproduces the accuracy of separately trained GNNs of each depth. Finally, the paper proposes a heuristic policy: rank nodes by a centrality measure (degree, k-core, PageRank, o

Load-bearing premise

The layer-selection rule assumes that nodes with similar centrality scores really do prefer the same exit layer, and that the layer chosen on validation nodes transfers to test nodes.

Editorial extensions

If this is right

  • Any message-passing GNN (the paper demonstrates GCN and GIN) can be wrapped with ADMP-GNN, and the paper shows both backbones improve on most of thirteen datasets.
  • Sequential training lets a single multi-exit network match the accuracy of separately trained per-depth GNNs, so adaptive depth does not require training L+1 independent models.
  • The oracle accuracy is far above the best fixed-depth accuracy, so per-node exit choice has headroom beyond what the current centrality policy harvests.
  • A fixed maximum depth of L=5 for ADMP-GNN competes with baselines whose depth is tuned per dataset, removing depth as a per-dataset tuning choice.
  • The paper reports no impactful accuracy drop from the adaptive scheme, meaning the added flexibility is not bought at the cost of worse predictions on any tested dataset.

Reading between the lines

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

  • The centrality-bucket policy is likely not the end of the road: a learned exit-prediction head or an optimal-stopping rule trained on validation nodes could close part of the gap between the policy accuracy and the oracle accuracy.
  • Because shallow-exit nodes skip computation, the same architecture could reduce inference cost in addition to improving accuracy, though the paper does not measure this; a direct FLOPs comparison across exit layers would test it.
  • The exit layers chosen per centrality bucket could serve as a cheap empirical probe of where oversmoothing or oversquashing sets in locally, connecting adaptive depth to known failure modes of deep GNNs.
  • Since the paper fixes the maximum depth for all datasets, the relative contribution of adaptive selection versus total depth is untested; varying both the maximum depth and the number of centrality buckets would separate the two effects.
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 / 4 minor

Summary. ADMP-GNN augments a message-passing GNN with a classifier ('exit') at every layer, trains the layers sequentially so that earlier layers are frozen ('ST'), and then learns a per-node exit-layer policy by clustering validation nodes according to centrality scores. The paper motivates the framework with synthetic experiments showing that sparse and dense subgraphs favor different depths, reports an oracle upper bound, and evaluates the approach on node classification benchmarks. The central claim is that dynamically adjusting the number of message-passing layers per node improves accuracy over fixed-depth GNNs.

Significance. If the central claim were established, ADMP-GNN would be a useful model-agnostic plug-in: one multi-exit model would match separately trained GNNs of each depth, and a cheap centrality heuristic would select a good exit per node. The sequential-training scheme itself is a genuine empirical contribution: Tables 1 and 8 show that ST largely avoids the collapse suffered by ALM and closely tracks single-task GCN/GIN at each depth. The authors also provide public code and an explicit, clearly defined oracle. However, the paper's headline performance claim is not supported by the reported policy results, and the experimental comparison contains several load-bearing weaknesses described below.

major comments (4)
  1. [§4.1–4.2, Tables 3/4] The rows labeled 'ADMP-GCN' and 'ADMP-GIN' in Tables 3/4 are not the adaptive policy: they are the best per-layer accuracies of the multi-exit model with the layer chosen on the test set (compare Table 2). The actual policy results are the 'w/Degree', 'w/k-core', 'w/Walk Count', and 'w/PageRank' rows. Under the actual policy, ADMP-GCN underperforms GCN on CS (e.g., 91.26–91.29 vs 92.20) and PubMed (78.64–78.71 vs 79.32), and ADMP-GIN underperforms GIN on CS and PubMed. Since baselines are test-set-tuned (Section 4.1) while the policy must transfer from validation, the comparison already favors baselines; yet no consistent gain emerges. The abstract's 'improved performance' and §4.2's 'consistently higher accuracy' claims are therefore not established. The authors should report the actual policy as the primary comparison, add significance tests, or substantially reframe the claim.
  2. [§3.4, Table 2] The oracle accuracy A_oracle is computed by using test labels to choose, for each test node, any layer whose prediction is correct. This is an upper bound, not evidence that a learnable policy can realize the adaptive benefit. The gap between the oracle and the actual policy rows is large (e.g., Cora 89.43 vs 81.19 for the best policy row in Table 3), showing that the centrality-based heuristic captures only a small fraction of the oracle's headroom. The oracle should not be used as support for the proposed policy unless accompanied by a comparable validation-based upper bound or an analysis of how much of the oracle gap is actually achievable without test labels.
  3. [§3.5] The layer-selection policy is underspecified and therefore not reproducible. The number C of centrality buckets is never reported, nor is the range or criterion used to choose it. The choice of centrality metric is dataset-dependent (§4.2 mentions k-core for Cora/ogbn-arxiv/Photo and Walk Count for Texas/Wisconsin), but no protocol is given for whether these choices were made using only validation data or with knowledge of test performance. Without these details, the 'w/...' results cannot be independently reproduced, and the claim that validation-chosen cluster exits transfer to test nodes is untested.
  4. [§4.1, Tables 3/4] The main comparison is incomplete and statistically unsupported. Although the paper states that thirteen datasets are used, Tables 3/4 report only six. Results for Chameleon, Squirrel, Computers, Photo, Cornell, Texas, and Wisconsin are referenced in the text/appendix but do not appear in the main experimental tables. Furthermore, no significance tests are reported, and many differences are within one standard deviation across the 10 runs. To support 'consistently higher' claims, the authors need to report all datasets and provide paired significance tests or effect sizes.
minor comments (4)
  1. [§3.3, Algorithm 1] In Algorithm 1, step 3 for t≥1 computes p^(t) using phi_Ct^(t-1), which is inconsistent with Eq. (3), where predictions are made by the exit update phi_Ex^(ℓ). Please correct the notation so the algorithm matches the architecture description.
  2. [Tables 3/5] Typos: Table 3 header spells 'Genuis' (should be 'Genius'); Table 5 spells 'squirel' and 'chamelon' (should be 'Squirrel' and 'Chameleon').
  3. [Appendix D] The dropout probability search space is listed as [16, 32, 64, 128, 256, 512], which is the same as the hidden-size search space and is not a valid dropout range. Please correct to actual dropout values.
  4. [Table 1] Several ST rows report a standard deviation of exactly 0.00 (e.g., Cora, layer 1; PubMed, layer 1). If 10 random seeds were used, this is surprising; please clarify whether the reported runs share initialization or whether the standard deviation is computed differently.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation; the adaptive-depth claim rests on standard train/validation/test separation, and the noted evaluation concerns are empirical rather than circular.

full rationale

The paper does not contain a derivation chain in which an output is equivalent to its inputs by construction. The main definitional object, Oracle Accuracy (Eq. in Section 3.4), is explicitly an upper bound computed from test labels, and it is not used to train any parameter or to select the final policy; it is only used to motivate the possibility of per-node exit layers. The layer-selection policy (Section 3.5) is learned from validation-node accuracy within centrality buckets and then applied to test nodes, which is a standard inductive step rather than a renaming of the training objective. The sequential training (ST) scheme is compared against single-task training and aggregate-loss training, and while the architecture is designed so that per-layer computational graphs match standard GNNs, the paper does not claim the equivalence as a theorem; it demonstrates it empirically in Tables 1 and 8. Self-citations (e.g., [17], [30], [31]) appear only as background references and are not load-bearing for the central claim. The most significant concern is that the rows labeled 'ADMP-GCN' and 'ADMP-GIN' in Tables 3/4 appear to be test-tuned best-single-layer results rather than the adaptive policy, and the actual policy rows ('w/Degree', 'w/k-core', etc.) do not consistently beat fixed-depth baselines. That is an empirical-support / evaluation-validity issue, not a circularity: the policy results are genuine predictions on held-out test nodes, and the best-layer rows are not disguised fits of the adaptive output. Accordingly, no specific circular step can be exhibited, and the circularity score is 0.

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

The central claim rests on two heuristic choices (centrality policy and multi-exit architecture) plus a fixed depth. No new physical or mathematical entities are introduced.

free parameters (3)
  • Cluster count C for centrality buckets
    The policy partitions nodes into C equal buckets; C controls policy granularity and is not given or ablated in the paper.
  • Centrality metric per dataset = e.g., k-core for Cora/ogbn-arxiv, Walk Count for Texas/Wisconsin
    The paper reports all four variants and suggests dataset-specific choices (Sec 4.2), but does not state whether the selection was made on validation or test data.
  • Maximum depth L = 5
    ADMP-GNN is evaluated with L fixed to 5, while baselines are tuned over 0..5 on the test set; this fixed choice is not justified.
assumptions (3)
  • domain assumption The exit classifier at layer l, which uses only aggregated neighbor messages m_v^(l), can match the accuracy of a standard GNN of depth l.
    Section 3.2 claims computational-graph equivalence to L+1 separately trained GNNs, but the exit ignores the node's own representation; only empirical support is provided in Table 1.
  • ad hoc to paper Optimal per-node depth is predictable from node centrality.
    Section 3.5 builds the entire test-time policy on this assumption; no evidence beyond anecdotal dataset observations is given.
  • domain assumption Sequential layer-by-layer training with freezing yields representations comparable to single-task training.
    Algorithm 1 and Section 3.3; empirically supported by Table 1 but not theoretically justified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of ADMP-GNN: Adaptive Depth Message Passing GNN." pith.science (2026). https://pith.science/paper/XOBX443H

@misc{pith2026250901170,
  author       = {Pith},
  title        = {Pith review of: ADMP-GNN: Adaptive Depth Message Passing GNN},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XOBX443H}},
  note         = {Machine review of arXiv:2509.01170}
}
read the original abstract

Graph Neural Networks (GNNs) have proven to be highly effective in various graph learning tasks. A key characteristic of GNNs is their use of a fixed number of message-passing steps for all nodes in the graph, regardless of each node's diverse computational needs and characteristics. Through empirical real-world data analysis, we demonstrate that the optimal number of message-passing layers varies for nodes with different characteristics. This finding is further supported by experiments conducted on synthetic datasets. To address this, we propose Adaptive Depth Message Passing GNN (ADMP-GNN), a novel framework that dynamically adjusts the number of message passing layers for each node, resulting in improved performance. This approach applies to any model that follows the message passing scheme. We evaluate ADMP-GNN on the node classification task and observe performance improvements over baseline GNN models.

Figures

Figures reproduced from arXiv: 2509.01170 by the authors.

Figure 1
Figure 1. Effect of GCN’s depth on sparse and dense sub [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Illustration of ADMP-GNN, when the maximum GNN depth is [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The adjacency matrix of the synthetic graphs ex [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

53 extracted references · 40 canonical work pages

  1. [1]

    Austin R Benson, David F Gleich, and Jure Leskovec. 2016. Higher-order organi- zation of complex networks. Science 353, 6295 (2016), 163–166

    work page 2016

  2. [2]

    Tolga Bolukbasi, Joseph Wang, Ofer Dekel, and Venkatesh Saligrama. 2017. Adap- tive neural networks for efficient inference. InInternational Conference on Machine Learning. PMLR, 527–536

    work page 2017

  3. [3]

    Stefan Bornholdt and Heinz Georg Schuster. 2001. Handbook of graphs and networks. From Genome to the Internet, Willey-VCH (2003 Weinheim) (2001)

    work page 2001

  4. [4]

    Sergey Brin and Lawrence Page. 1998. The anatomy of a large-scale hypertextual web search engine. Computer networks and ISDN systems 30, 1-7 (1998), 107–117

    work page 1998

  5. [5]

    Defu Cao, Yujing Wang, Juanyong Duan, Ce Zhang, Xia Zhu, Congrui Huang, Yunhai Tong, Bixiong Xu, Jing Bai, Jie Tong, et al. 2020. Spectral temporal graph neural network for multivariate time-series forecasting. Advances in neural information processing systems 33 (2020), 17766–17778

    work page 2020

  6. [6]

    Castro-Correa, Jhony H

    Jhon A. Castro-Correa, Jhony H. Giraldo, Mohsen Badiey, and Fragkiskos D. Malliaros. 2024. Gegenbauer Graph Neural Networks for Time-Varying Signal Reconstruction. IEEE Transactions on Neural Networks and Learning Systems 35, 9 (2024), 11734–11745

    work page 2024

  7. [7]

    Ming Chen, Zhewei Wei, Zengfeng Huang, Bolin Ding, and Yaliang Li. 2020. Simple and deep graph convolutional networks. In International conference on machine learning. PMLR, 1725–1735

    work page 2020

  8. [8]

    Eli Chien, Jianhao Peng, Pan Li, and Olgica Milenkovic. 2020. Adaptive universal generalized pagerank graph neural network. arXiv preprint arXiv:2006.07988 (2020)

    arXiv 2020

Show all 53 references
  1. [9]

    Gabriele Corso, Bowen Jing, Regina Barzilay, Tommi Jaakkola, et al. 2023. Diff- Dock: Diffusion Steps, Twists, and Turns for Molecular Docking. In International Conference on Learning Representations (ICLR 2023)

    2023

  2. [10]

    Malliaros, Yoshua Bengio, and David Rolnick

    Alexandre Duval, Victor Schmidt, Alex Hernández-García, Santiago Miret, Fragkiskos D. Malliaros, Yoshua Bengio, and David Rolnick. 2023. FAENet: Frame Averaging Equivariant GNN for Materials Modeling. In Proceedings of the 40th International Conference on Machine Learning . PM...

    2023

  3. [11]

    Maha Elbayad, Jiatao Gu, Edouard Grave, and Michael Auli. 2020. Depth-Adaptive Transformer. In International Conference on Learning Representations . https: //openreview.net/forum?id=SJg7KhVKPH

    2020

  4. [12]

    Moshe Eliasof, Beatrice Bevilacqua, Carola-Bibiane Schönlieb, and Haggai Maron

  5. [13]

    Federico Errica, Henrik Christiansen, Viktor Zaverkin, Takashi Maruyama, Math- ias Niepert, and Francesco Alesiani. 2023. Adaptive Message Passing: A General Framework to Mitigate Oversmoothing, Oversquashing, and Underreaching. arXiv preprint arXiv:2312.16560 (2023)

    2023 arXiv

  6. [14]

    Lukas Faber and Roger Wattenhofer. 2024. GwAC: GNNs with Asynchronous Communication. In Learning on Graphs Conference . PMLR, 8–1

    2024

  7. [15]

    Bronstein, and Ismail Ilkan Ceylan

    Ben Finkelshtein, Xingyue Huang, Michael M. Bronstein, and Ismail Ilkan Ceylan

  8. [16]

    Justin Gilmer, Samuel S Schoenholz, Patrick F Riley, Oriol Vinyals, and George E Dahl. 2017. Neural message passing for quantum chemistry. In International conference on machine learning . PMLR, 1263–1272

    2017

  9. [17]

    https://openreview.net/forum?id= T0FuEDnODP

    Cooperative Graph Neural Networks. https://openreview.net/forum?id= T0FuEDnODP

  10. [18]

    Alex Graves. 2016. Adaptive computation time for recurrent neural networks. arXiv preprint arXiv:1603.08983 (2016)

    2016 arXiv

  11. [19]

    Giraldo, Konstantinos Skianis, Thierry Bouwmans, and Fragkiskos D

    Jhony H. Giraldo, Konstantinos Skianis, Thierry Bouwmans, and Fragkiskos D. Malliaros. 2023. On the Trade-off between Over-smoothing and Over-squashing in Deep Graph Neural Networks. In Proceedings of the 32nd ACM International Conference on Information and Knowledge Managemen...

    2023

  12. [20]

    Gao Huang, Yu Sun, Zhuang Liu, Daniel Sedra, and Kilian Q Weinberger. 2016. Deep networks with stochastic depth. In Computer Vision–ECCV 2016: 14th Euro- pean Conference, Amsterdam, The Netherlands, October 11–14, 2016, Proceedings, CIKM ’25, November 10–14, 2025, Seoul, Repub...

    2016

  13. [21]

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

    2020

  14. [22]

    Xu Jia, Bert De Brabandere, Tinne Tuytelaars, and Luc V Gool. 2016. Dynamic filter networks. Advances in neural information processing systems 29 (2016)

    2016

  15. [23]

    Côme Huré, Huyên Pham, Achref Bachouch, and Nicolas Langrené. 2021. Deep neural networks algorithms for stochastic control problems on finite horizon: convergence analysis. SIAM J. Numer. Anal. 59, 1 (2021), 525–557

    2021

  16. [24]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In ICLR

    2017

  17. [25]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. 2014. Adam: A Method for Stochastic Opti- mization. doi:10.48550/ARXIV.1412.6980

    2014 doi arXiv

  18. [26]

    Derek Lim and Austin R Benson. 2021. Expertise and dynamics within crowd- sourced musical knowledge curation: A case study of the genius platform. In Proceedings of the International AAAI Conference on Web and Social Media , Vol. 15. 373–384

    2021

  19. [27]

    Hao Li, Asim Kadav, Igor Durdanovic, Hanan Samet, and Hans Peter Graf. 2017. Pruning Filters for Efficient ConvNets. In International Conference on Learning Representations. https://openreview.net/forum?id=rJqFGTslg

    2017

  20. [28]

    Juncheng Liu, Kenji Kawaguchi, Bryan Hooi, Yiwei Wang, and Xiaokui Xiao

  21. [29]

    Derek Lim, Felix Matthew Hohne, Xiuyu Li, Sijia Linda Huang, Vaishnavi Gupta, Omkar Prasad Bhalerao, and Ser-Nam Lim. 2021. Large Scale Learning on Non- Homophilous Graphs: New Benchmarks and Strong Simple Methods. InAdvances in Neural Information Processing Systems , A. Beyge...

    2021

  22. [30]

    Malliaros, Christos Giatsidis, Apostolos N

    Fragkiskos D. Malliaros, Christos Giatsidis, Apostolos N. Papadopoulos, and Michalis Vazirgiannis. 2020. The core decomposition of networks: theory, algo- rithms and applications. VLDB J. 29, 1 (2020), 61–92

    2020

  23. [31]

    Malliaros and Michalis Vazirgiannis

    Fragkiskos D. Malliaros and Michalis Vazirgiannis. 2013. Clustering and com- munity detection in directed networks: A survey. Physics Reports 533, 4 (2013), 95–142. Clustering and Community Detection in Directed Networks: A Survey

    2013

  24. [32]

    Sitao Luan, Chenqing Hua, Qincheng Lu, Jiaqi Zhu, Mingde Zhao, Shuyuan Zhang, Xiao-Wen Chang, and Doina Precup. 2022. Revisiting heterophily for graph neural networks. Advances in neural information processing systems 35 (2022), 1362–1375

    2022

  25. [33]

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

    2022

  26. [34]

    Benedek Rozemberczki, Carl Allen, and Rik Sarkar. 2021. Multi-scale attributed node embedding. Journal of Complex Networks 9, 2 (2021), cnab014

    2021

  27. [35]

    George Panagopoulos, Nikolaos Tziortziotis, Michalis Vazirgiannis, and Fragkiskos Malliaros. 2024. Maximizing Influence with Graph Neural Networks. In Proceedings of the 2023 IEEE/ACM International Conference on Advances in Social Networks Analysis and Mining . 237–244

    2024

  28. [36]

    Tran, Yi Tay, and Donald Metzler

    Tal Schuster, Adam Fisch, Jai Gupta, Mostafa Dehghani, Dara Bahri, Vinh Q. Tran, Yi Tay, and Donald Metzler. 2022. Confident Adaptive Language Modeling. In Advances in Neural Information Processing Systems , Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho (Eds...

    2022

  29. [37]

    Prithviraj Sen, Galileo Namata, Mustafa Bilgic, Lise Getoor, Brian Galligher, and Tina Eliassi-Rad. 2008. Collective Classification in Network Data. AI Magazine 29, 3 (Sep. 2008), 93. doi:10.1609/aimag.v29i3.2157

    2008 doi

  30. [38]

    Sara Sabour, Nicholas Frosst, and Geoffrey E Hinton. 2017. Dynamic routing between capsules. Advances in neural information processing systems 30 (2017)

    2017

  31. [39]

    Indro Spinelli, Simone Scardapane, and Aurelio Uncini. 2020. Adaptive propa- gation graph convolutional network. IEEE Transactions on Neural Networks and Learning Systems 32, 10 (2020), 4755–4760

    2020

  32. [40]

    Ke Sun, Zhanxing Zhu, and Zhouchen Lin. 2021. Ada{GCN}: Adaboosting Graph Convolutional Networks into Deep Models. In International Conference on Learn- ing Representations. https://openreview.net/forum?id=QkRbdiiEjM

    2021

  33. [41]

    Oleksandr Shchur, Maximilian Mumme, Aleksandar Bojchevski, and Stephan Günnemann. 2018. Pitfalls of Graph Neural Network Evaluation. In NeurIPS Relational Representation Learning Workshop (R2L 2018)

    2018

  34. [42]

    Yulin Wang, Kangchen Lv, Rui Huang, Shiji Song, Le Yang, and Gao Huang

  35. [43]

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. 2019. How Pow- erful are Graph Neural Networks?. In 7th International Conference on Learning Representations

    2019

  36. [44]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. Attention Is All You Need. Advances in neural information processing systems (2017)

    2017

  37. [45]

    Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. 2018. Representation learning on graphs with jumping knowledge networks. In International conference on machine learn- ing. pmlr, 5453–5462

    2018

  38. [46]

    Hanqing Zeng, Muhan Zhang, Yinglong Xia, Ajitesh Srivastava, Rajgopal Kannan, Viktor Prasanna, Long Jin, Andrey Malevich, and Ren Chen. 2020. Deep graph neural networks with shallow subgraph samplers. (2020)

    2020

  39. [47]

    Wentao Zhang, Zeang Sheng, Yuezihan Jiang, Yikuan Xia, Jun Gao, Zhi Yang, and Bin Cui. 2021. Evaluating deep graph neural networks. arXiv preprint arXiv:2108.00955 (2021)

    2021 arXiv

  40. [48]

    Keyulu Xu, Chengtao Li, Yonglong Tian, Tomohiro Sonobe, Ken-ichi Kawarabayashi, and Stefanie Jegelka. 2018. Representation Learning on Graphs with Jumping Knowledge Networks. CoRR abs/1806.03536 (2018). arXiv:1806.03536 http://arxiv.org/abs/1806.03536

    2018 arXiv

  41. [49]

    McAuley, Ke Xu, and Furu Wei

    Wangchunshu Zhou, Canwen Xu, Tao Ge, Julian J. McAuley, Ke Xu, and Furu Wei. 2020. BERT Loses Patience: Fast and Robust Inference with Early Exit.CoRR abs/2006.04152 (2020). arXiv:2006.04152 https://arxiv.org/abs/2006.04152

    2020 arXiv

  42. [52]

    Lingxiao Zhao and Leman Akoglu. 2020. PairNorm: Tackling Oversmoothing in GNNs. In International Conference on Learning Representations

    2020

  43. [2020]

    CoRR abs/2010.05300 (2020)

    Glance and Focus: a Dynamic Approach to Reducing Spatial Redundancy in Image Classification. CoRR abs/2010.05300 (2020). arXiv:2010.05300 https: //arxiv.org/abs/2010.05300

    2010 arXiv

  44. [2021]

    Advances in Neural Information Processing Systems 34 (2021), 18762–18773

    Eignn: Efficient infinite-depth graph neural networks. Advances in Neural Information Processing Systems 34 (2021), 18762–18773

    2021

  45. [2024]

    arXiv preprint arXiv:2404.13344 (2024)

    GRANOLA: Adaptive Normalization for Graph Neural Networks. arXiv preprint arXiv:2404.13344 (2024)

    2024 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.