Pith. sign in

REVIEW 5 major objections 8 minor 52 references

Robust Learning on Noisy Graphs via Latent Space Constraints with External Knowledge

T0 review · 5 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read LSC-GNN claims that adding a penalty between embeddings from a noisy full graph and a clean external graph stops graph neural networks from overfitting spurious edges, and shows gains over standard and noise-resilient baselines on…

desk verdict A plausible dual-encoder regularization idea that works in several benchmark settings, but the paper never pins down whether the clean reference stays clean, and a definitional slip muddies the method. read the letter →

arxiv 2507.05540 v1 pith:D2MNN5LC submitted 2025-07-07 cs.LG cs.AI

Chunhui Gu , Mohammad Sadegh Nasr , James P. Long , Kim-Anh Do , Ehsan Irajizad This is my paper · ORCID
classification cs.LGcs.AI MSC 68T0768R10
keywords graphneuralnetworksnoisyedgeslatentspaceregularizationexternalknowledgenodeclassificationheterogeneousgraphsprotein-metabolitenetworkrobustness
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

The paper introduces LSC-GNN, a training method for Graph Neural Networks that makes node classification robust to noisy edges by leaning on cleaner external links. The idea is to train two encoders at once: one on the full graph (target plus external edges) and another on a regularization graph that excludes the target's potentially spurious edges, then add a penalty for disagreement between their latent embeddings. The claim is that this constraint steers the main encoder away from overfitting false connections while preserving genuine structure. Experiments on Cora, CiteSeer, and PubMed with injected noise show LSC-GNN matching or beating standard GNNs and dedicated noise-resilient baselines at moderate perturbation rates, and a small protein–metabolite case study lifts ROC-AUC from 0.92 to 0.96 when curated metabolite–protein links are added with the constraint. If correct, this offers a simple, backbone-agnostic way to inject trusted external knowledge into noisy graph learning.

What carries the argument

The central mechanism is dual-encoder latent-space regularization: two message-passing GNN encoders with the same architecture, one operating on the full graph and one on the clean regularization graph, with an $\ell^2$ penalty on the difference between their node embeddings. The regularization graph is built by removing all target-graph edges (the potentially noisy ones) and keeping only external edges and their nodes, so the second encoder sees only trusted structure. The penalty term $L_{\mathrm{reg}} = \frac{1}{|V_t|L}\sum_{k\in V_t}\sum_{m=1}^L (z_{k,m} - z'_{k,m})^2$ is what transfers the clean graph's geometry into the noisy graph's learning, under the stated assumption that true meaningful links form tightly connected clusters in the clean larger graph while noisy links do not.

What would settle it

Injected-noise experiments on the external graph itself: take a benchmark graph, add false positive edges to the regularization graph $G_r$ at rates comparable to the target noise, and test whether LSC-GNN's accuracy drops to or below that of a single-encoder GNN trained on the noisy full graph alone. A stronger falsifier is to use a completely unrelated external graph (no informative overlap) and show the method cannot beat the no-regularization baseline, demonstrating that the reported gains come entirely from external graph quality.

Watch

Extended reading notes

Core claim

LSC-GNN establishes that noisy-edge robustness can come from a soft constraint between two views of the same nodes. Given a noisy target graph $G_t$ and a cleaner external graph $G_r$ sharing nodes with it, the full graph $G_f = G_t \cup G_r$ is fed to a main GNN encoder $f$, while $G_r$ alone is fed to a second encoder $f'$. The total loss is $L_{\mathrm{total}} = L_{\mathrm{target}} + \lambda L_{\mathrm{reg}}$, where $L_{\mathrm{reg}}$ is the mean squared difference between the embeddings of target nodes produced by $f$ and $f'$ (Equation 6). This penalizes the main encoder precisely when it learns representations unsupported by clean external structure, preventing overfit to spurious edges. The authors show that with only about 10% of external links, the approach outperforms GCN, GAT, ProGNN, GCN-Jaccard, and GCNSVD on benchmark citation graphs under low-to-moderate random perturbation, and they extend the same dual-encoder constraint to heterogeneous graphs where the regularization graph is formed by dropping the target edge type.

Load-bearing premise

The external graph must actually be clean, informative, and overlapping enough with the target nodes; if the external links are themselves noisy, incomplete, or irrelevant, the regularization penalty will push the learned embeddings away from the true structure rather than toward it.

Editorial extensions

If this is right

  • LSC-GNN provides a drop-in regularization that works with any message-passing encoder (the paper uses GAT) and does not require modifying the adjacency matrix or the main architecture.
  • The method is effective even when only a small fraction (about 10%) of external links are available, making it practical for settings where trusted auxiliary data is scarce.
  • The same latent-space constraint extends to heterogeneous graphs with multiple edge types, opening a route to multi-omics and knowledge-graph applications without new architectural machinery.
  • Because regularization strength is selected on the validation set, the authors argue the method is guaranteed not to hurt performance compared with ignoring the external graph entirely.
  • The framework naturally fits settings where a target graph is a subgraph of a larger, better-curated network, such as biological interaction networks, social graphs, or citation networks.

Reading between the lines

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

  • A testable extension the authors do not run: perturb the external graph itself (inject false edges or drop informative links) and measure when LSC-GNN's benefit disappears—this would calibrate how clean the 'clean' graph must be.
  • The dense-cluster assumption suggests the method's gain should scale with the assortativity or community structure of the external graph; varying that structure could predict which real datasets benefit.
  • The dual-encoder penalty could be adapted to semi-supervised or link-prediction losses beyond node classification, since the regularization operates only on embeddings, not on the task head.
  • In practice, this is effectively a form of contrastive or view-consistency learning where the 'positive view' is provided by trusted relational data, so the method could be combined with other view-perturbation augmentations.
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 / 8 minor

Summary. The paper proposes Latent Space Constrained Graph Neural Networks (LSC-GNN), a dual-encoder framework that trains one GNN on a full graph (noisy target edges plus clean external edges) and a second GNN on a regularization graph that excludes the target edges, then penalizes the mean squared difference between the two latent spaces. The authors claim this steers the full-graph encoder away from spurious edges, and they report experiments on Cora, CiteSeer, and PubMed under random false-positive edge perturbation, plus a small heterogeneous protein–metabolite case study. The paper also includes an ablation study, complexity analysis, and a discussion of limitations.

Significance. If the regularization mechanism works as claimed, LSC-GNN would provide a simple, backbone-agnostic way to inject trusted external link structure into GNN training for noisy graphs, which is a practically relevant and underexplored direction. The paper is explicitly non-circular in its evaluation: the regularization graph is an intended input, the coefficient is chosen on validation, and test results are external. The release of code and the use of standard benchmarks are strengths. However, the central mechanism is not yet established: the formal definition of the regularization graph conflicts with the loss, the regularization encoder has no independent anchor and may drift, the empirical advantage is not supported by significance tests, and the heterogeneous claim rests on a single un-replicated case study. These issues are addressable but are load-bearing for the paper's main claims.

major comments (5)
  1. [Section 3.3, Eq. (6), Algorithm 1] Problem 1 defines Gr = (Vr, Er, Xf) with Vr = Vf \ Vt, so the regularization graph excludes the target nodes. Yet Eq. (6) computes Lreg over k ∈ Vt using z^r_k, the embedding of target node k from f'(Gr), which is undefined as written. Algorithm 1 and Section 4.1.3 instead construct Gr by removing the target edges while retaining target nodes and the cross edges between target and external nodes. This inconsistency must be fixed: either the formal definition of Gr should include the target nodes and external-to-target edges, or Eq. (6) must be restricted to nodes that actually appear in Gr. The choice matters because the presence or absence of target nodes in Gr changes the information available to the regularization encoder.
  2. [Section 4.1.2, Eq. (5)-(6)] The central claim that the regularization term 'steers' the full-graph encoder away from spurious edges is not established because f' has no independent anchor. As written, Lreg is the only training signal for f': f' receives no task loss, no reconstruction loss on Gr, and no stop-gradient, and gradients flow symmetrically into both f and f'. Nothing prevents f' from gradually adapting to reproduce f's noisy embeddings, in which case Lreg becomes small even when f overfits spurious edges. The ablation in Section 6 compares LSC-GNN against single-graph models but does not isolate whether f' maintains a genuinely clean reference; it only shows that the combined training objective helps on the studied setting. Please provide evidence about the training dynamics (e.g., measure the distance between f'(Gr) and the embedding of a fixed clean-reference encoder), or modify the architecture to anchor f' (e.g., stop-gradient on the regularization branch, pretrain f' on a reconstruction task over Gr, or freeze f' after an initial phase).
  3. [Section 11] The statement that selecting the best regularization strength on validation 'thereby guaranteeing no detrimental effect on the final results' is logically incorrect. Model selection on the validation set does not guarantee test performance; a configuration that looks better on validation can still generalize worse on the held-out test set. The 'incremental regularization strengths' used for this selection are also not described in Section 5. Please remove the guarantee wording and replace it with an empirical claim supported by the reported experiments, or provide a formal argument for why validation-based selection of λ transfers to the test set in this setting.
  4. [Table 2 and Section 5.1] The conclusion that LSC-GNN 'outperforms' baselines is not supported by statistical significance testing. Many reported margins are within one standard deviation of the baselines (for example, Cora ratio 0.50 at noise 0.25: LSCGNN 76.63±1.28 vs. GCN-Jaccard 74.59±2.11; PubMed ratio 0.50 at noise 0.10: LSCGNN 82.9±1.07 vs. GCNSVD 84.38±0.14). In addition, at zero perturbation LSC-GNN is sometimes worse than baselines (Cora ratio 0.50: LSCGNN 78.0±1.47 vs. GCN 79.11±1.59), which is consistent with a regularization penalty but should be explicitly discussed. Please report paired significance tests over the 10 seeds (e.g., Wilcoxon signed-rank test) and adjust the strength of the claims accordingly.
  5. [Section 9.4] The heterogeneous extension is validated on a single small protein–metabolite network with single ROC-AUC numbers (0.92, 0.94, 0.96), no error bars, no repeated-seed statistics, and no comparison to a heterogeneous GNN baseline other than the full-graph variant of the same method. Given that 'extension to heterogeneous graphs' is listed as one of the three main contributions, this evidence is too thin to support the claim. At minimum, report mean±std over multiple random seeds and compare against a standard heterogeneous GNN (e.g., RGCN or HGT) trained on the full graph without the regularization branch.
minor comments (8)
  1. [Abstract and Section 1] There are several grammatical slips, e.g., 'in graph subjected to moderate noise' should be 'in graphs subjected to moderate noise', and 'can improved accuracy' in Section 10 should be 'can improve accuracy'.
  2. [Section 3.2] The text says 'in Kipf(2013)', but the surrounding reference list cites Kipf and Welling's variational graph auto-encoder paper from 2016; the intended citation should be corrected and made consistent.
  3. [Eq. (3)] The softmax notation is inconsistent: the equation uses softmax_j(e_ij) but then the denominator is written as a sum over k, and the indices in the sum should be checked (the sum should be over neighbors of i, not over k with inconsistent notation).
  4. [Section 4.1.2] 'ensuing comparability' should be 'ensuring comparability', and the attention coefficient formula in Eq. (2) uses both a^k and W^k without defining their dimensions or the initialization.
  5. [Section 7 and Figure 2] The runtime figure is not referenced or explained in the text; please specify what is being measured, on which graph scale, and under what hardware configuration, so that the complexity analysis in the same section can be interpreted.
  6. [Section 9.2] 'GraphSAEG' appears to be a typo for 'GraphSAGE'; please correct it and check for similar spelling errors throughout the manuscript.
  7. [Section 9.3] The sentence 'The dataset is split into 50% train, 20% validation, and 30% test simulate a real-world scenario' is missing the word 'to' before 'simulate'.
  8. [Section 5] The paper says each hyperparameter configuration was evaluated with 10 random seeds, but it is not stated whether the node splits and the false-positive edge sampling are fixed or re-sampled across seeds; please clarify what variability the error bars capture.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LSC-GNN's external regularization graph is a genuine input, hyperparameters are selected on validation, and test metrics are held out; the main caveats are soundness issues rather than circular reductions.

full rationale

No circularity is established. The regularization graph Gr is defined as external data in Problem 1: "Define the regularization graph Gr = (Vr, Er, Xf), which isolates only the external edges and their associated nodes." It is not a fitted quantity, a renamed prediction, or a function of the target labels. The target loss is computed from target labels, Lreg is a discrepancy penalty between two encoders, and the regularization coefficient is selected on a validation set: "the model with best validation performance was evaluated on the test dataset." Test ROC-AUC values are therefore external to the training objective. The main claim that the constraint "steers the model away from overfitting spurious edges" is an empirical claim supported by comparisons against single-graph and full-graph variants in Section 6; it is not true by construction, but it is also not a prediction that reduces to its input. The paper has no load-bearing self-citations and imports no uniqueness theorem to force its design. The skeptic's concerns about the anchoring mechanism, namely that f' is trained only by the same MSE and that Eq. (6) indexes target nodes not present in Gr as defined in Problem 1, are validity and reproducibility concerns about whether the clean reference is guaranteed to stay clean; they are not instances of a derived result being equivalent to its inputs by definition. The Limitations section's statement that validation-based selection "guarantees no detrimental effect on the final results" is an overclaim, but it concerns model selection and generalizability, not circularity. Accordingly, the paper's derivation chain is self-contained with respect to the circularity patterns considered.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The method has one clear free hyperparameter, lambda, tuned on validation. The main axiom is that the external graph is clean and relevant, which is stated but not justified or tested. The MSE latent-space penalty is a design choice with no supporting analysis. No new entities are postulated.

free parameters (1)
  • regularization coefficient lambda = tuned per dataset on validation set
    Equation (5) weighs the latent-space regularization term; the paper states the model with best validation performance is selected, so lambda is effectively a free hyperparameter.
assumptions (3)
  • domain assumption External edges in the regularization graph are clean and informative about target nodes
    Invoked in Section 4.1.1 and Problem 1: the method relies on Gr being a reliable signal. Acknowledged as a limitation in Section 11.
  • domain assumption Original benchmark graph edges are true links, so edges removed from the target graph and kept in the regularization graph are clean
    In the simulation (Section 4.1.3), the original graph is treated as ground truth, and the regularization graph is built from those assumed-clean edges.
  • ad hoc to paper Mean squared error between latent representations is a meaningful regularization signal
    Equation (6) defines Lreg as MSE between the two encoders' outputs; no theoretical or empirical justification is given for why this particular discrepancy measure aligns embeddings with external knowledge beyond the stated assumption.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Learning on Noisy Graphs via Latent Space Constraints with External Knowledge." pith.science (2026). https://pith.science/paper/D2MNN5LC

@misc{pith2026250705540,
  author       = {Pith},
  title        = {Pith review of: Robust Learning on Noisy Graphs via Latent Space Constraints with External Knowledge},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D2MNN5LC}},
  note         = {Machine review of arXiv:2507.05540}
}
read the original abstract

Graph Neural Networks (GNNs) often struggle with noisy edges. We propose Latent Space Constrained Graph Neural Networks (LSC-GNN) to incorporate external "clean" links and guide embeddings of a noisy target graph. We train two encoders--one on the full graph (target plus external edges) and another on a regularization graph excluding the target's potentially noisy links--then penalize discrepancies between their latent representations. This constraint steers the model away from overfitting spurious edges. Experiments on benchmark datasets show LSC-GNN outperforms standard and noise-resilient GNNs in graphs subjected to moderate noise. We extend LSC-GNN to heterogeneous graphs and validate it on a small protein-metabolite network, where metabolite-protein interactions reduce noise in protein co-occurrence data. Our results highlight LSC-GNN's potential to boost predictive performance and interpretability in settings with noisy relational structures.

Figures

Figures reproduced from arXiv: 2507.05540 by the authors.

Figure 1
Figure 1. Framework of LSC-GNN: (1) The target graph contains noisy links (dashed lines), while [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 3
Figure 3. Ablation analysis. 8 Expansion to heterogeneous graph Although the framework described thus far focuses on homogeneous graphs, where all nodes and edges share a single modality, many real-world applications—such as multi-omics biological net￾works, knowledge graphs, and recommendation systems—feature multiple node and edge types. The latent space regularization strategy introduced above can be naturally adapted to t… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

52 extracted references · 42 canonical work pages

  1. [1]

    Modeling polypharmacy side effects with graph convolutional networks

    Marinka Zitnik, Monica Agrawal, and Jure Leskovec. Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics, 34(13):i457–i466, June 2018

    work page 2018

  2. [2]

    Graph neural networks for social recommendation, 2019

    Wenqi Fan, Yao Ma, Qing Li, Yuan He, Eric Zhao, Jiliang Tang, and Dawei Yin. Graph neural networks for social recommendation, 2019

    2019

  3. [3]

    Neural relational inference for interacting systems, 2018

    Thomas Kipf, Ethan Fetaya, Kuan-Chieh Wang, Max Welling, and Richard Zemel. Neural relational inference for interacting systems, 2018

    work page 2018

  4. [4]

    Variational graph auto-encoders

    Thomas N Kipf and Max Welling. Variational graph auto-encoders. arXiv preprint arXiv:1611.07308, 2016

    arXiv 2016

  5. [5]

    NED-GNN: Detecting and Dropping Noisy Edges in Graph Neural Networks , pages 91–105

    Ming Xu, Baoming Zhang, Jinliang Yuan, Meng Cao, and Chongjun Wang. NED-GNN: Detecting and Dropping Noisy Edges in Graph Neural Networks , pages 91–105. Springer Nature Switzerland, 2023

    work page 2023

  6. [6]

    Towards robust graph neural networks for noisy graphs with sparse labels, 2022

    Enyan Dai, Wei Jin, Hui Liu, and Suhang Wang. Towards robust graph neural networks for noisy graphs with sparse labels, 2022

    work page 2022

  7. [7]

    Lichtenwalter, and Nitesh V

    Yang Yang, Ryan N. Lichtenwalter, and Nitesh V . Chawla. Evaluating link prediction methods. Knowledge and Information Systems, 45(3):751–782, 2015

    work page 2015

  8. [8]

    Evaluating graph neural networks for link prediction: Current pitfalls and new benchmarking, 2023

    Juanhui Li, Harry Shomer, Haitao Mao, Shenglai Zeng, Yao Ma, Neil Shah, Jiliang Tang, and Dawei Yin. Evaluating graph neural networks for link prediction: Current pitfalls and new benchmarking, 2023

    work page 2023

Show all 52 references
  1. [9]

    Graph convolution for semi- supervised classification: Improved linear separability and out-of-distribution generalization, 2022

    Aseem Baranwal, Kimon Fountoulakis, and Aukosh Jagannath. Graph convolution for semi- supervised classification: Improved linear separability and out-of-distribution generalization, 2022

    2022

  2. [10]

    Uncovering disease-disease relationships through the incomplete interactome

    Jörg Menche, Amitabh Sharma, Maksim Kitsak, Susan Dina Ghiassian, Marc Vidal, Joseph Loscalzo, and Albert-László Barabási. Uncovering disease-disease relationships through the incomplete interactome. Science, 347(6224):1257601, 2015

    2015

  3. [11]

    Ned-gnn: Detecting and dropping noisy edges in graph neural networks

    Ming Xu, Baoming Zhang, Jinliang Yuan, Meng Cao, and Chongjun Wang. Ned-gnn: Detecting and dropping noisy edges in graph neural networks. In Asia-Pacific Web (APWeb) and Web-Age Information Management (WAIM) Joint International Conference on Web and Big Data, pages 91–105. Spr...

    2022

  4. [12]

    Adver- sarial examples on graph data: Deep insights into attack and defense, 2019

    Huijun Wu, Chen Wang, Yuriy Tyshetskiy, Andrew Docherty, Kai Lu, and Liming Zhu. Adver- sarial examples on graph data: Deep insights into attack and defense, 2019

    2019

  5. [13]

    Garnet: Reduced-rank topology learning for robust and scalable graph neural networks

    Chenhui Deng, Xiuyu Li, Zhuo Feng, and Zhiru Zhang. Garnet: Reduced-rank topology learning for robust and scalable graph neural networks. In Learning on Graphs Conference, pages 3–1. PMLR, 2022. 10

    2022

  6. [14]

    All you need is low (rank) defending against adversarial attacks on graphs

    Negin Entezari, Saba A Al-Sayouri, Amirali Darvishzadeh, and Evangelos E Papalexakis. All you need is low (rank) defending against adversarial attacks on graphs. In Proceedings of the 13th international conference on web search and data mining, pages 169–177, 2020

    2020

  7. [15]

    Elastic graph neural networks

    Xiaorui Liu, Wei Jin, Yao Ma, Yaxin Li, Hua Liu, Yiqi Wang, Ming Yan, and Jiliang Tang. Elastic graph neural networks. In Marina Meila and Tong Zhang, editors, Proceedings of the 38th International Conference on Machine Learning, volume 139 of Proceedings of Machine Learning R...

    2021

  8. [16]

    Gnnguard: Defending graph neural networks against adversarial attacks

    Xiang Zhang and Marinka Zitnik. Gnnguard: Defending graph neural networks against adversarial attacks. Advances in neural information processing systems, 33:9263–9275, 2020

    2020

  9. [17]

    Ad- versarial examples on graph data: Deep insights into attack and defense

    Huijun Wu, Chen Wang, Yuriy Tyshetskiy, Andrew Docherty, Kai Lu, and Liming Zhu. Ad- versarial examples on graph data: Deep insights into attack and defense. arXiv preprint arXiv:1903.01610, 2019

    1903 arXiv

  10. [18]

    Graphdefense: Towards robust graph convolutional networks

    Xiaoyun Wang, Xuanqing Liu, and Cho-Jui Hsieh. Graphdefense: Towards robust graph convolutional networks. arXiv preprint arXiv:1911.04429, 2019

    1911 arXiv

  11. [19]

    Topology attack and defense for graph neural networks: An optimization perspective

    Kaidi Xu, Hongge Chen, Sijia Liu, Pin-Yu Chen, Tsui-Wei Weng, Mingyi Hong, and Xue Lin. Topology attack and defense for graph neural networks: An optimization perspective. arXiv preprint arXiv:1906.04214, 2019

    1906 arXiv

  12. [20]

    Graph structure learning for robust graph neural networks, 2020

    Wei Jin, Yao Ma, Xiaorui Liu, Xianfeng Tang, Suhang Wang, and Jiliang Tang. Graph structure learning for robust graph neural networks, 2020

    2020

  13. [21]

    Transferring robustness for graph neural network against poisoning attacks

    Xianfeng Tang, Yandong Li, Yiwei Sun, Huaxiu Yao, Prasenjit Mitra, and Suhang Wang. Transferring robustness for graph neural network against poisoning attacks. In Proceedings of the 13th International Conference on Web Search and Data Mining, WSDM ’20. ACM, January 2020

    2020

  14. [22]

    Variational inference for graph convolutional networks in the absence of graph data and adversarial settings

    Pantelis Elinas, Edwin V Bonilla, and Louis Tiao. Variational inference for graph convolutional networks in the absence of graph data and adversarial settings. Advances in neural information processing systems, 33:18648–18660, 2020

    2020

  15. [23]

    Robustness of graph neural networks at scale

    Simon Geisler, Tobias Schmidt, Hakan Sirin, Daniel Zügner, Aleksandar Bojchevski, and Stephan Günnemann. Robustness of graph neural networks at scale. Advances in Neural Information Processing Systems, 34:7637–7649, 2021

    2021

  16. [24]

    Graph attention networks, 2018

    Petar Velickovic, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Liò, and Yoshua Bengio. Graph attention networks, 2018

    2018

  17. [25]

    How powerful are graph neural networks?, 2019

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks?, 2019

    2019

  18. [26]

    Collective classification in network data

    Prithviraj Sen, Galileo Mark Namata, Mustafa Bilgic, Lise Getoor, Brian Gallagher, and Tina Eliassi-Rad. Collective classification in network data. AI Magazine, 29(3):93–106, 2008

    2008

  19. [27]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization, 2017

    2017

  20. [28]

    Graph attention networks

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

    2017 arXiv

  21. [29]

    Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling

    Michael Schlichtkrull, Thomas N. Kipf, Peter Bloem, Rianne van den Berg, Ivan Titov, and Max Welling. Modeling relational data with graph convolutional networks, 2017

    2017

  22. [30]

    Heterogeneous graph transformer, 2020

    Ziniu Hu, Yuxiao Dong, Kuansan Wang, and Yizhou Sun. Heterogeneous graph transformer, 2020

    2020

  23. [31]

    Kegg: kyoto encyclopedia of genes and genomes

    Minoru Kanehisa and Susumu Goto. Kegg: kyoto encyclopedia of genes and genomes. Nucleic acids research, 28(1):27–30, 2000

    2000

  24. [32]

    Proteinbert: a universal deep-learning model of protein sequence and function

    Nadav Brandes, Dan Ofer, Yam Peleg, Nadav Rappoport, and Michal Linial. Proteinbert: a universal deep-learning model of protein sequence and function. Bioinformatics, 38(8):2102– 2110, 2022. 11

    2022

  25. [33]

    Rdkit: A software suite for cheminformatics, computational chemistry, and predictive modeling.[google scholar]

    G Landrum. Rdkit: A software suite for cheminformatics, computational chemistry, and predictive modeling.[google scholar]. 2013

    2013

  26. [34]

    Lu, Kevin K

    Christian Dallago, Konstantin Schütze, Michael Heinzinger, Tobias Olenyi, Maria Littmann, Amy X. Lu, Kevin K. Yang, Seonwoo Min, Sungroh Yoon, James T. Morton, and Burkhard Rost. Learned embeddings from deep learning to visualize and predict protein sets. Current Protocols, 1(...

    2021

  27. [35]

    Mpi-vgae: protein–metabolite enzymatic reaction link learning by variational graph autoencoders

    Cheng Wang, Chuang Yuan, Yahui Wang, Ranran Chen, Yuying Shi, Tao Zhang, Fuzhong Xue, Gary J Patti, Leyi Wei, and Qingzhen Hou. Mpi-vgae: protein–metabolite enzymatic reaction link learning by variational graph autoencoders. Briefings in bioinformatics, 24(4):bbad189, 2023

    2023

  28. [36]

    Matthias Fey and Jan E. Lenssen. Fast graph representation learning with PyTorch Geometric. In ICLR Workshop on Representation Learning on Graphs and Manifolds, 2019. A Technical Appendices and Supplementary Material We use IBM Load Sharing Facility (LSF) system in Linux to ex...

    2019

  29. [37]

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: The main claims in the abstract and introduction are well-aligned with the paper’s contributions and scope. The introd...

  30. [38]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We discuss the limitation of approach in the conclusion section. The approach is limited to the existence of a high-quality external graph that are c...

  31. [39]

    • All the theorems, formulas, and proofs in the paper should be numbered and cross- referenced

    Theory assumptions and proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [NA] Justification: [NA] Guidelines: • The answer NA means that the paper does not include theoretical results. •...

  32. [40]

    Experimental result reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  33. [41]

    The main page of the repository contains a detailed instruction to replicate all results in this paper

    Open access to data and code Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] 15 Justification: We have attached the Git...

  34. [42]

    Experimental setting/details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: We clearly specifies the trai...

  35. [43]

    We explicitly state that each hyper-parameter configuration is run with 10 different random initializations on a fixed node splits

    Experiment statistical significance 16 Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: For every experimental setting we reports mean ...

  36. [44]

    Experiments compute resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: We have provided the CP...

  37. [45]

    • If the authors answer No, they should explain the special circumstances that require a deviation from the Code of Ethics

    Code of ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: Guidelines: • The answer NA means that the authors have not reviewed the NeurIP...

  38. [46]

    And we also include a potential negative societal impact in the appendix

    Broader impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: We have discuss the potential positive societal impacts in biology field in the discussion section. And w...

  39. [47]

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: Guidelines: • T...

  40. [48]

    • The authors should cite the original paper that produced the code package or dataset

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: Guidelines: • Th...

  41. [49]

    New assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [NA] Justification: Datasets: Cora, CiteSeer, and PubMed datasets are used and cited appropriately ([26]), which are publicly available b...

  42. [50]

    Crowdsourcing and research with human subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  43. [51]

    • Depending on the country in which research is conducted, IRB approval (or equivalent) may be required for any human subjects research

    Institutional review board (IRB) approvals or equivalent for research with human subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

  44. [52]

    Declaration of LLM usage Question: Does the paper describe the usage of LLMs if it is an important, original, or non-standard component of the core methods in this research? Note that if the LLM is used only for writing, editing, or formatting purposes and does not impact the ...

    2025

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.