Pith. sign in

REVIEW 3 major objections 6 minor 299 references

On the same graph, node-label supervision reliably helps link prediction on homophilic graphs, but link supervision helps node classification only in a structure-dominant regime—and only under a leakage-free protocol.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 04:37 UTC pith:5NPM77HF

load-bearing objection Clean same-graph NC–LP protocol with a real directional finding; soft spot is the collection-relative “structure-dominant” label, not the asymmetry itself. the 3 major comments →

arxiv 2607.28525 v1 pith:5NPM77HF submitted 2026-07-30 cs.LG cs.SI

Same Graph Cross-Task Transfer in GNNs: Protocols and Predictors

Neelam Akula , Surbhi Kumar , Murat Kantarcioglu , Baris Coskunuzer This is my paper
classification cs.LG cs.SI
keywords graph neural networksnode classificationlink predictioncross-task transferhomophilymulti-task learningevaluation protocolCoTask Score
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Real graphs often carry both node labels and observed edges, so practitioners want to reuse one task’s supervision to help the other, or train one shared encoder for both. Prior comparisons mixed incompatible splits, message-passing graphs, and negative samples, so reported transfer gains were untrustworthy. This paper fixes a single leakage-free protocol—shared node and edge splits, message passing only on training edges, and fixed link-prediction negatives—and measures bidirectional transfer across three backbone families. It finds a stable directional pattern: node classification improves link prediction on homophilic graphs, while the reverse direction is fragile and can hurt accuracy unless link prediction is already easy and node classification still has headroom. Homophily and simple task-learnability statistics then guide which transfer mechanism to use and when a shared encoder is worth the risk.

Core claim

Under a standardized leakage-free same-graph protocol, NC→LP transfer is consistently beneficial on homophilic graphs across GCN, GraphSAGE, and GPS, whereas LP→NC is unreliable under naive representation reuse and becomes reliably positive mainly in a structure-dominant regime where link prediction is highly learnable but node classification remains unsaturated. Coupled training (multi-view and joint) is generally more stable than post-hoc reuse, and the CoTask Score summarizes when one shared encoder serves both tasks without net loss.

What carries the argument

Leakage-free evaluation protocol: fixed NC node splits and LP edge splits, a single observed adjacency built only from training positive edges for all message passing, and fixed LP negatives reused across methods. CoTask Score (CTS) averages percent gains on NC accuracy and LP AUC over a fixed GCN reference to score joint utility.

Load-bearing premise

The “structure-dominant” regime is defined relative to ranking within this fixed collection of eleven datasets; if the pool or backbone ranking changes, which graphs count as structure-dominant—and thus where reverse transfer is called reliable—can shift.

What would settle it

Rerun the same protocol on a larger, independently chosen set of graphs with the same three backbones: if NC→LP gains no longer track homophily, or if LP→NC gains appear strongly outside the high-LP / low-NC-headroom quadrant (or vanish inside it), the directional regime claims fail.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Default to NC→LP reuse on high-homophily graphs; treat LP→NC as optional structural pretraining only when LP is easy and NC is unsaturated.
  • Prefer multi-view or joint training over embedding replacement when a single encoder must serve both tasks, to reduce negative transfer.
  • Report CTS alongside per-task metrics so shared-encoder choices are judged by net NC+LP utility, not one task alone.
  • Use edge/node homophily (and baseline task learnability) as cheap pre-checks before committing to a transfer direction or mechanism.
  • Future multi-objective graph benchmarks should adopt fixed splits, observed graphs, and negatives so cross-task claims are comparable.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same protocol could be reused as a stress test for whether newer long-range or state-space graph models preserve the NC→LP / LP→NC asymmetry or erase it.
  • If structure-dominant reverse transfer is real, cheap unsupervised link objectives may substitute for scarce node labels precisely on graphs where connectivity is informative but labels are not aligned with edges.
  • Collection-relative regime labels suggest publishing absolute baseline LP/NC scores with every claim, so later work can re-bin regimes without re-running the original suite.
  • Negative transfer under ET-Replace implies frozen LP embeddings often destroy class-separating geometry; adaptive gating of transferred features may be a practical next control knob.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. The paper formalizes same-graph cross-task transfer between node classification (NC) and link prediction (LP), and introduces a leakage-free evaluation protocol that fixes node/edge splits, restricts message passing to A_obs built only from training positives, and reuses fixed LP negatives across methods. Under this protocol it evaluates five lightweight transfer regimes (WS, ET-Rep, ET-Concat, MV, Joint) on three backbones (GCN, GraphSAGE, GPS) over 11 datasets spanning homophilic, heterophilic, and mixed regimes, in both transductive and (partially) inductive settings. The central empirical claim is a directional, regime-dependent asymmetry: NC→LP is consistently beneficial on homophilic graphs and correlates with homophily, whereas LP→NC is fragile under naive representation reuse and is reliably positive mainly in a structure-dominant regime (high LP learnability, unsaturated NC). The authors further define CoTask Score (CTS) to summarize joint NC+LP utility for shared-encoder deployments and argue that simple diagnostics (especially homophily) can guide mechanism choice and avoid negative transfer.

Significance. If the findings hold, the work supplies a practical and overdue evaluation standard for multi-objective GNN pipelines that currently mix incompatible splits, observed graphs, and negative-sampling rules. The directional asymmetry and the homophily–NC→LP correlation are actionable for practitioners choosing whether and how to reuse supervision, and CTS is a simple, reproducible multi-objective summary. Strengths that raise confidence include: explicit leakage sanity checks; consistency across three architecturally distinct backbones; supporting inductive tables; a second hyperparameter configuration (A.6); two extra homophilic Amazon graphs (A.7); and geometric evidence via cosine-similarity gaps (Table 11, Fig. 2). Code and fixed-split reproducibility are promised. The contribution is primarily empirical and protocol-level rather than architectural, which is appropriate given the stated goal.

major comments (3)
  1. [Section 5, Datasets; also Abstract and §5.1] Section 5 (Datasets): the operational definition of the structure-dominant regime—baseline LP AUC in the upper half and baseline NC accuracy in the lower half of this fixed 11-dataset collection under a given backbone—is collection-relative. The claim that LP→NC is “reliably positive mainly” in that regime (Abstract, §5.1, Conclusion) therefore depends on membership of USA/EUROPE/BRAZIL within this pool. Please either (i) state an a priori, collection-independent criterion (e.g., absolute LP AUC and NC headroom thresholds, or a leave-one-dataset-out membership check), or (ii) rephrase the claim as an empirical concentration on these three graphs and report sensitivity of membership and of the LP→NC conclusion when the pool or backbone ranking changes. Without that, the regime language overclaims relative to the evidence.
  2. [Appendix A.4, Tables 9–10; §5.1] Tables 9–10 (Appendix A.4): inductive results for MV and Joint are reported only under GCN, while GraphSAGE and GPS stop at WS/ET. The paper’s robustness claim for coupled objectives (MV/Joint as the “safest” choices in §5.1–5.2) is therefore only partially stress-tested under distribution shift to unseen nodes. Either complete MV/Joint for the other backbones in the inductive setting, or clearly scope the inductive conclusions to WS/ET and avoid generalizing the coupling recommendation from inductive GCN alone.
  3. [Section 3.1 Eq. (1); Table 5; Table 8] Section 3.1, Eq. (1) and Table 5: CTS is anchored to a fixed GCN reference g for all backbones. This is fine for within-GCN comparison, but when GraphSAGE/GPS rows are scored against a GCN base, large positive CTS on heterophilic graphs can reflect backbone superiority rather than transfer benefit (e.g., GraphSAGE/GPS Base CTS already large on TEXAS/WISCONSIN). Please either (i) also report backbone-local CTS anchored to each backbone’s own single-task base, or (ii) separate “transfer gain CTS” from “backbone CTS” in Table 5/8 so joint-utility claims about transfer mechanisms are not confounded by encoder choice.
minor comments (6)
  1. [Appendix A.1] Appendix A.1, Eq. (2): λ_x is initialized from a homophily/CC formula before validation search. You correctly note that final λ_x is validation-selected and that asymmetry appears in non-MV methods; still, state explicitly in the main text (not only appendix) that MV results are not driven by this schedule, and report the selected λ_x values per dataset.
  2. [Table 3 caption] Table 3 caption says gains are “relative to the corresponding LP single-task base model”; it should say NC single-task base. Same check for related captions.
  3. [Table 1; Section 5] Table 1: USA/EUROPE/BRAZIL list “–” for features; clarify in text whether one-hot or structural features are used and whether that interacts with ET-Rep/Concat.
  4. [Figure 1; §4] Figure 1 is conceptually helpful but the “Warm Start Path” and multiple minima are schematic only; a short caption caveat that this is intuition, not a computed loss landscape, would avoid over-reading.
  5. [Appendix A.2] Typos/style: “Formal defintions” (A.2); “defintions”; occasional missing spaces before citations; arXiv ID formatting in the header is fine for preprint but journal version should drop the duplicate abstract block artifacts if any remain.
  6. [Section 2] Related Work: a brief pointer to recent multi-task GNN recommendation / joint NC–LP production systems would better situate the deployment motivation already stated in the introduction.

Circularity Check

1 steps flagged

No load-bearing circular derivation; only mild post-hoc labeling of the structure-dominant regime from the same collection’s baseline ranks.

specific steps
  1. other [Section 5, Datasets (structure-dominant definition)]
    "under a fixed backbone and protocol, a dataset is structure-dominant if its baseline LP AUC ranks in the upper half of the benchmark while its baseline NC accuracy ranks in the lower half. ... Empirically, USA, EUROPE, and BRAZIL satisfy this criterion across all three backbones under our protocol"

    Regime membership is fixed from the same 11-dataset collection’s baseline rank halves, then the narrative claim that LP→NC is “reliably positive mainly” in that regime is read off the same collection. This is mild post-hoc labeling, not identity-by-construction: transfer gains are still independently measured relative to NC bases, and the definition uses baselines rather than the LP→NC outcomes themselves. It does not force the directional asymmetry or the homophily–NC→LP correlations.

full rationale

The paper’s central claims are empirical comparisons under a fixed leakage-free protocol: bidirectional transfer gains versus single-task bases (Tables 2–3, 9–10), homophily–gain correlations (Tables 4, 7), CTS as an explicit relative summary (Eq. 1), and geometric cosine-gap evidence (Table 11). None of these reduce by construction to their inputs. CTS and S_val are openly normalized reporting/selection metrics, not disguised first-principles predictions. The sole mild concern is that “structure-dominant” is defined from within-collection baseline LP/NC rank halves and then used to locate where LP→NC gains concentrate; that is post-hoc regime naming on the same benchmark pool, not a fitted parameter renamed as a prediction, and membership is fixed from baselines rather than from the transfer outcomes themselves. No self-citation uniqueness chain, ansatz smuggling, or self-definitional identity between claimed outputs and defining inputs appears. Score 1 reflects that minor labeling issue only.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 3 invented entities

The work is empirical ML: claims rest on standard GNN training assumptions, a hand-specified evaluation protocol, and operational regime labels—not on new physical entities. Load-bearing choices are split ratios, fixed 1:1 negatives, A_obs = training positives only, CTS anchored to a GCN reference, and the half-rank definition of structure-dominant datasets.

free parameters (6)
  • NC label split ratios 60%/20%/20% = 60/20/20
    Chosen as a fixed protocol constant across all experiments; not derived; affects supervision available for NC and transfer.
  • LP positive edge split ratios 80%/10%/10% = 80/10/10
    Standard but free protocol choice held fixed; changes LP train signal and A_obs density.
  • LP negative sampling ratio 1:1 with fixed seed policy = |E-_s|=|E+_s|
    Defines the LP classification task difficulty; fixed across methods but still a design parameter.
  • Joint task weight λ and MV alignment weight λ_x = λ∈{0.0..1.0} grid; λ_x∈[0.2,1.6]
    Tuned on validation (λ grid; λ_x in [0.2,1.6], with homophily/CC-based initialization formula then validated). Final values are data-selected hyperparameters that affect coupled-regime outcomes.
  • CTS reference model g = standard GCN under protocol = GCN single-task reference
    CTS percent gains are anchored to this fixed reference per dataset; choice of g rescales the joint score used for multi-objective comparison.
  • Default encoder hyperparameters (depth, width, dropout, lr, patience) = e.g. 2-layer, hid 64, dropout 0.5, lr 0.01, patience 50
    Tuned once on validation and shared across regimes; sensitivity rerun uses an alternate config, showing patterns hold but absolute gains shift.
axioms (6)
  • ad hoc to paper Message-passing embeddings for all tasks/methods must be computed only on A_obs := Adj(V, E+_tr), excluding val/test LP positives.
    Core protocol axiom (Section 3) defining leakage-free same-graph evaluation; standard motivation from LP literature but the unified NC+LP interface is paper-specified.
  • ad hoc to paper LP evaluation uses fixed negative sets E-_s generated once and reused across methods and runs.
    Required so method comparisons do not confound transfer with resampling (Section 3).
  • domain assumption Homophily (edge/node) is a meaningful diagnostic of label-connectivity alignment for predicting NC→LP transfer.
    Standard graph ML assumption (Luan et al., Platonov et al.); used as predictor in Tables 4 and 7.
  • ad hoc to paper A dataset is structure-dominant if baseline LP AUC ranks upper half and baseline NC accuracy ranks lower half of the fixed 11-dataset collection under a given backbone/protocol.
    Operational regime definition in Section 5; load-bearing for the LP→NC 'mainly in structure-dominant regime' claim.
  • domain assumption GCN, GraphSAGE, and GPS are representative enough that directional findings are not backbone artifacts within message-passing and graph-transformer families.
    Stated scope limitation in Related Work/Conclusion; authors do not claim universality to state-space or other new architectures.
  • domain assumption Standard supervised losses (cross-entropy NC; BCE LP) and MLP pairwise decoder over [zu∥zv∥|zu−zv|∥zu⊙zv] adequately probe representation transfer.
    Common LP head design (Section 4); results could in principle depend on decoder choice.
invented entities (3)
  • CoTask Score (CTS) no independent evidence
    purpose: Scalar summary of joint NC+LP utility as average percent gain over a fixed GCN reference on accuracy and AUC.
    Defined in Section 3.1 Eq. (1); reporting/selection metric, not a physical entity. Independent meaning only relative to chosen reference g and task metrics.
  • Structure-dominant regime (operational half-rank definition) no independent evidence
    purpose: Label datasets where LP→NC is expected to act as structural pretraining (easy LP, unsaturated NC).
    Introduced to organize LP→NC successes on USA/EUROPE/BRAZIL; membership is benchmark-relative, so external falsifiability requires freezing the collection and protocol.
  • Leakage-free same-graph NC–LP evaluation protocol independent evidence
    purpose: Standardize splits, A_obs, and fixed negatives for bidirectional transfer comparison.
    Primary methodological contribution (Section 3); reusable checklist rather than a latent scientific object.

pith-pipeline@v1.2.0-daily-grok45 · 29147 in / 4320 out tokens · 86548 ms · 2026-07-31T04:37:49.188968+00:00 · methodology

0 comments
read the original abstract

Many real-world graphs support multiple predictive tasks over the same underlying structure, creating an opportunity to reuse supervision across node classification (NC) and link prediction (LP). However, existing evaluations often rely on incompatible splits, observed-graph assumptions, and negative sampling rules, making conclusions about same-graph cross-task transfer unreliable. We formalize same-graph NC-LP transfer and propose a leakage-free protocol that fixes node and edge splits, uses a shared message-passing graph that excludes evaluated edges, and employs fixed negatives for LP. Across three backbones (GCN, GraphSAGE, GPS), we find that transfer is strongly directional and predictable: NC $\to$ LP is consistently beneficial on homophilic graphs, while LP $\to$ NC is fragile and can even degrade accuracy under naive representation reuse. LP $\to$ NC becomes reliably positive mainly in a structure-dominant regime where LP is easy but NC is unsaturated, suggesting that LP acts as structural pretraining. Finally, we introduce the CoTask Score (CTS) to summarize joint NC+LP utility when a shared encoder must serve both tasks, and show that simple dataset statistics, especially homophily, can guide mechanism choice and help avoid negative transfer.

Figures

Figures reproduced from arXiv: 2607.28525 by Baris Coskunuzer, Murat Kantarcioglu, Neelam Akula, Surbhi Kumar.

Figure 1
Figure 1. Figure 1: Cross-task transfer as optimization in a shared pa￾rameter space. For a fixed GNN/GT architecture, NC and LP correspond to different losses over the same parameters, yielding distinct (possibly incompatible) minima. Warm start initializes the target training near a source-task solution, while MV and Joint couple objectives to steer optimization toward solutions that per￾form well on both tasks; other local… view at source ↗
Figure 2
Figure 2. Figure 2: NC training amplifies cosine similarity gaps on homophilic graphs. Left: The increase in cosine similarity gap between linked and unlinked pairs after NC training (versus raw features) correlates strongly with node homophily (r = +0.901∗∗), explaining why NC→LP transfer is reliably beneficial on homophilic datasets. Right: On homophilic graphs, NC-trained embeddings (orange) produce substantially larger al… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

299 extracted references · 41 linked inside Pith

  1. [1]

    AAAI , volume=

    Hypergraph Neural Architecture Search , author=. AAAI , volume=

  2. [2]

    arXiv preprint arXiv:2409.02901 , year=

    Topological Methods in Machine Learning: A Tutorial for Practitioners , author=. arXiv preprint arXiv:2409.02901 , year=

    Pith/arXiv arXiv

  3. [3]

    2008 , publisher=

    Introduction to information retrieval , author=. 2008 , publisher=

    2008

  4. [4]

    The Web Conference , pages=

    Graph contrastive learning with adaptive augmentation , author=. The Web Conference , pages=

  5. [5]

    NeurIPS , year=

    From canonical correlation analysis to self-supervised graph neural networks , author=. NeurIPS , year=

  6. [6]

    Physica A: Statistical Mechanics and its Applications , volume =

    Costa, Luciano da Fontoura , title =. Physica A: Statistical Mechanics and its Applications , volume =. 2022 , doi =

    2022

  7. [7]

    TMLR , year=

    Single-Pass Contrastive Learning Can Work for Both Homophilic and Heterophilic Graph , author=. TMLR , year=

  8. [8]

    NeurIPS , volume=

    Decoupled self-supervised learning for graphs , author=. NeurIPS , volume=

  9. [9]

    Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

    Towards self-supervised learning on graphs with heterophily , author=. Proceedings of the 31st ACM International Conference on Information & Knowledge Management , pages=

  10. [10]

    arXiv preprint arXiv:2212.04604 , year=

    Localized Contrastive Learning on Graphs , author=. arXiv preprint arXiv:2212.04604 , year=

    Pith/arXiv arXiv

  11. [11]

    International Conference on Learning Representations , year=

    Large-scale representation learning on graphs via bootstrapping , author=. International Conference on Learning Representations , year=

  12. [12]

    NeurIPS , volume=

    Simple and asymmetric graph contrastive learning without augmentations , author=. NeurIPS , volume=

  13. [13]

    IEEE Transactions on Knowledge and Data Engineering , year=

    Make Heterophilic Graphs Better Fit GNN: A Graph Rewiring Approach , author=. IEEE Transactions on Knowledge and Data Engineering , year=

  14. [14]

    AAAI , volume=

    Beyond smoothing: Unsupervised graph representation learning with edge heterophily discriminating , author=. AAAI , volume=

  15. [15]

    IEEE Data Engineering Bulletin , year=

    Heterophily and graph neural networks: Past, present and future , author=. IEEE Data Engineering Bulletin , year=

  16. [16]

    NeurIPS , volume=

    From trainable negative depth to edge heterophily in graphs , author=. NeurIPS , volume=

  17. [17]

    Learning on Graphs Conference , pages=

    Edge directionality improves learning on heterophilic graphs , author=. Learning on Graphs Conference , pages=. 2024 , organization=

    2024

  18. [18]

    NeurIPS , volume=

    Demystifying structural disparity in graph neural networks: Can one size fit all? , author=. NeurIPS , volume=

  19. [19]

    ACM Transactions on Intelligent Systems and Technology (TIST) , volume=

    Graph neural networks: Taxonomy, advances, and trends , author=. ACM Transactions on Intelligent Systems and Technology (TIST) , volume=. 2022 , publisher=

    2022

  20. [20]

    NeurIPS , volume=

    Diverse message passing for attribute with heterophily , author=. NeurIPS , volume=

  21. [21]

    NeurIPS , volume=

    Bernnet: Learning arbitrary graph spectral filters via bernstein approximation , author=. NeurIPS , volume=

  22. [22]

    Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

    Ud-gnn: Uncertainty-aware debiased training on semi-homophilous graphs , author=. Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining , pages=

  23. [23]

    Proceedings of the ACM Web Conference 2022 , pages=

    Gbk-gnn: Gated bi-kernel graph neural networks for modeling both homophily and heterophily , author=. Proceedings of the ACM Web Conference 2022 , pages=

    2022

  24. [24]

    2022 IEEE International Conference on Data Mining (ICDM) , pages=

    Hp-gmn: Graph memory networks for heterophilous graphs , author=. 2022 IEEE International Conference on Data Mining (ICDM) , pages=. 2022 , organization=

    2022

  25. [25]

    , author=

    Visualizing data using t-SNE. , author=. JMLR , volume=

  26. [26]

    Forty-first International Conference on Machine Learning , year=

    Position: Graph Foundation Models Are Already Here , author=. Forty-first International Conference on Machine Learning , year=

  27. [27]

    Liu, Juncheng and Hooi, Bryan and Kawaguchi, Kenji and Xiao, Xiaokui , journal=

  28. [28]

    IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

    A survey of vectorization methods in topological data analysis , author=. IEEE Transactions on Pattern Analysis and Machine Intelligence , year=

  29. [29]

    Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , pages=

    MUSE: Multi-View Contrastive Learning for Heterophilic Graphs , author=. Proceedings of the 32nd ACM International Conference on Information and Knowledge Management , pages=

  30. [30]

    The Eighth International Conference on Learning Representations (ICLR 2020), Addis Ababa , year=

    Measuring and improving the use of graph information in graph neural network , author=. The Eighth International Conference on Learning Representations (ICLR 2020), Addis Ababa , year=

    2020

  31. [31]

    arXiv preprint arXiv:2006.07988 , year=

    Adaptive universal generalized pagerank graph neural network , author=. arXiv preprint arXiv:2006.07988 , year=

    Pith/arXiv arXiv 2006

  32. [32]

    AAAI , volume=

    Beyond low-frequency information in graph convolutional networks , author=. AAAI , volume=

  33. [33]

    NeurIPS , volume=

    Inductive representation learning on large graphs , author=. NeurIPS , volume=

  34. [34]

    KDD , pages=

    Structural deep network embedding , author=. KDD , pages=

  35. [35]

    Proceedings of the 24th international conference on world wide web , pages=

    Line: Large-scale information network embedding , author=. Proceedings of the 24th international conference on world wide web , pages=

  36. [36]

    Less is More: SlimG for Accurate, Robust, and Interpretable Graph Mining , author=

  37. [37]

    arXiv preprint arXiv:2109.05641 , year=

    Is heterophily a real nightmare for graph neural networks to do node classification? , author=. arXiv preprint arXiv:2109.05641 , year=

    Pith/arXiv arXiv

  38. [38]

    Revisiting heterophily for graph neural networks , author=

  39. [39]

    When do graph neural networks help with node classification: Investigating the homophily principle on node distinguishability , author=

  40. [40]

    Deep Gaussian Embedding of Graphs: Unsupervised Inductive Learning via Ranking , author=

  41. [41]

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

  42. [42]

    2023 , organization=

    On over-squashing in message passing neural networks: The impact of width, depth, and topology , author=. 2023 , organization=

    2023

  43. [43]

    What graph neural networks cannot learn: depth vs width , author=

  44. [44]

    Uncertainty in Artificial Intelligence , pages=

    Studying the Effect of GNN Spatial Convolutions On The Embedding Space’s Geometry , author=. Uncertainty in Artificial Intelligence , pages=. 2023 , organization=

    2023

  45. [45]

    IEEE Transactions on Neural Networks and Learning Systems , volume=

    A comprehensive survey on graph neural networks , author=. IEEE Transactions on Neural Networks and Learning Systems , volume=. 2020 , publisher=

    2020

  46. [46]

    How hard is to distinguish graphs with graph neural networks? , author=

  47. [47]

    arXiv preprint arXiv:2003.04078 , year=

    A survey on the expressive power of graph neural networks , author=. arXiv preprint arXiv:2003.04078 , year=

    Pith/arXiv arXiv 2003

  48. [48]

    Approximate graph propagation , author=

  49. [49]

    Am-gcn: Adaptive multi-channel graph convolutional networks , author=

  50. [50]

    Diffusion improves graph learning , author=

  51. [51]

    Hu, Weihua and others , year=

  52. [52]

    KDD , pages=

    node2vec: Scalable feature learning for networks , author=. KDD , pages=

  53. [53]

    Proceedings of the 24th ACM International on Conference on Information and Knowledge Management , pages =

    Cao, Shaosheng and Lu, Wei and Xu, Qiongkai , title =. Proceedings of the 24th ACM International on Conference on Information and Knowledge Management , pages =. 2015 , isbn =. doi:10.1145/2806416.2806512 , abstract =

    arXiv 2015

  54. [54]

    Deepwalk: Online learning of social representations , author=

  55. [55]

    IEEE Transactions on Signal Processing , volume=

    Cayleynets: Graph convolutional neural networks with complex rational spectral filters , author=. IEEE Transactions on Signal Processing , volume=. 2018 , publisher=

    2018

  56. [56]

    arXiv preprint arXiv:2007.16002 , year=

    Graph convolutional networks using heat kernel for semi-supervised learning , author=. arXiv preprint arXiv:2007.16002 , year=

    Pith/arXiv arXiv 2007

  57. [57]

    Proceedings of the 19th

    Diffusion kernels on graphs and other discrete structures , author=. Proceedings of the 19th

  58. [58]

    ICLR , year=

    Predict then Propagate: Graph Neural Networks meet Personalized PageRank , author=. ICLR , year=

  59. [59]

    , author=

    The PageRank citation ranking: Bringing order to the web. , author=. 1999 , institution=

    1999

  60. [60]

    AAAI , volume=

    Deeper insights into graph convolutional networks for semi-supervised learning , author=. AAAI , volume=

  61. [61]

    arXiv preprint arXiv:1412.6550 , year=

    Fitnets: Hints for thin deep nets , author=. arXiv preprint arXiv:1412.6550 , year=

    Pith/arXiv arXiv

  62. [62]

    arXiv preprint arXiv:1503.02531 , year=

    Distilling the knowledge in a neural network , author=. arXiv preprint arXiv:1503.02531 , year=

    Pith/arXiv arXiv

  63. [63]

    2022 , issue_date =

    Li, Xue and Cheng, Yuanzhi , title =. 2022 , issue_date =. doi:10.1016/j.knosys.2022.109919 , journal =

    arXiv 2022

  64. [64]

    The Eleventh

    ContraNorm: A Contrastive Learning Perspective on Oversmoothing and Beyond , author=. The Eleventh. 2023 , url=

    2023

  65. [65]

    arXiv preprint arXiv:2012.15024 , year=

    Adaptive Graph Diffusion Networks , author=. arXiv preprint arXiv:2012.15024 , year=

    Pith/arXiv arXiv 2012

  66. [66]

    arXiv preprint arXiv:2210.14709 , year=

    Learning on Large-scale Text-attributed Graphs via Variational Inference , author=. arXiv preprint arXiv:2210.14709 , year=

    Pith/arXiv arXiv

  67. [67]

    arXiv preprint arXiv:2304.10668 , year=

    Train Your Own GNN Teacher: Graph-Aware Distillation on Textual Graphs , author=. arXiv preprint arXiv:2304.10668 , year=

    Pith/arXiv arXiv

  68. [68]

    arXiv preprint arXiv:2204.01855 , year=

    A survey on graph representation learning methods , author=. arXiv preprint arXiv:2204.01855 , year=

    Pith/arXiv arXiv

  69. [69]

    2016 , note =

    Benchmark Data Sets for Graph Kernels , author =. 2016 , note =

    2016

  70. [70]

    Journal of Symbolic Computation , volume=

    A persistence landscapes toolbox for topological statistics , author=. Journal of Symbolic Computation , volume=. 2017 , publisher=

    2017

  71. [71]

    how graph neural networks go beyond weisfeiler-lehman?

    A new perspective on" how graph neural networks go beyond weisfeiler-lehman?" , author=

  72. [72]

    AAAI , volume=

    Weisfeiler and leman go neural: Higher-order graph neural networks , author=. AAAI , volume=

  73. [73]

    Graph Neural Networks Exponentially Lose Expressive Power for Node Classification , author=

  74. [74]

    Information Sciences , volume=

    IEA-GNN: Anchor-aware graph neural network fused with information entropy for node classification and link prediction , author=. Information Sciences , volume=. 2023 , publisher=

    2023

  75. [75]

    How Powerful are Graph Neural Networks? , author=

  76. [76]

    ICLR , year=

    Wasserstein Embedding for Graph Learning , author=. ICLR , year=

  77. [77]

    Foundations of Computational Mathematics , pages=

    Homological algebra for persistence modules , author=. Foundations of Computational Mathematics , pages=. 2021 , publisher=

    2021

  78. [78]

    Graph Neural Networks: Foundations, Frontiers, and Applications , pages=

    Graph neural networks: Adversarial robustness , author=. Graph Neural Networks: Foundations, Frontiers, and Applications , pages=. 2022 , publisher=

    2022

  79. [79]

    International Journal of Bifurcation and Chaos , volume=

    Centrality estimation in large networks , author=. International Journal of Bifurcation and Chaos , volume=. 2007 , publisher=

    2007

  80. [80]

    Journal of mathematical sociology , volume=

    A faster algorithm for betweenness centrality , author=. Journal of mathematical sociology , volume=. 2001 , publisher=

    2001

Showing first 80 references.