Pith. sign in

REVIEW 2 major objections 7 minor 124 references

A nonlinear Laplacian for signed-directed graphs keeps only direction-aligned edges, is proved positive semidefinite with eigenvalues in [0,2], and powers a spectral GNN that outperforms prior models on node classification and link predicti

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 · deepseek-v4-flash

2026-08-05 00:12 UTC pith:ZDIN7YUU

load-bearing objection A genuinely new and provably sound nonlinear Laplacian for signed-directed graphs, attached to a GNN whose empirical claims outrun what the paper actually shows. the 2 major comments →

arxiv 2608.00836 v1 pith:ZDIN7YUU submitted 2026-08-01 cs.LG

Nonlinear Laplacians Improve Signed-Directed Graph Learning

Ali Parviz , Yuichi Yoshida This is my paper
classification cs.LG MSC 05C5005C2005C22
keywords signed-directed graphsnonlinear Laplacianspectral graph neural networkslink predictionnode classificationsigned graphsdirected graphsgraph convolution
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.

The paper tries to establish that signed and directed graphs—networks whose edges carry a positive/negative sign and a one-way direction—can be handled by a single nonlinear Laplacian operator, and that a spectral graph neural network built on this operator outperforms existing GNNs on node classification and link prediction. The operator works by assigning each node a scalar potential and keeping an edge only when the tail's potential is at least the head's potential; the kept edges form a signed undirected graph whose ordinary signed Laplacian is applied to the potentials. The paper proves this operator is positive semidefinite, with normalized eigenvalues between 0 and 2, and that its quadratic form has a clear combinatorial meaning: it penalizes misoriented positive edges and rewards cutting negative edges. If correct, this gives GNN designers a parameter-free spectral operator that integrates sign and direction simultaneously, which earlier Laplacians did not do.

Core claim

The central claim is Definition 2: for a signed-directed graph G=(V,E+,E-,w) and any vector x of node potentials, build a signed undirected graph G_x by taking each directed edge (u,v) of either sign and keeping it as an undirected edge of the same sign with weight w(e) when x_u >= x_v, and otherwise replacing it with self-loops of weight w(e)/2 on u and v. Then L_G(x) is defined as the signed-undirected Laplacian L_{G_x} x. The paper shows its quadratic form is the sum over positive edges of w(e) max{x_u - x_v,0}^2 plus the sum over negative edges of w(e)(2x_u^2 + 2x_v^2 - max{x_u - x_v,0}^2), so a negative edge contributes (x_u + x_v)^2 when x_u >= x_v and 2x_u^2 + 2x_v^2 otherwise. For a

What carries the argument

The load-bearing object is the nonlinear signed-directed Laplacian L_G(x) of Definition 2. It is built in three steps: (i) all directed edges whose direction agrees with the ordering of the potentials—tail x_u at least as large as head x_v—are kept as undirected edges of the same sign; (ii) every non-aligned edge is replaced by self-loops on its endpoints; (iii) the resulting signed undirected graph G_x is given to the classical signed Laplacian, so L_G(x) = L_{G_x} x. The quadratic-form identity connecting this operator to cut structure is what gives it combinatorial meaning, and the spectral bounds [0,2] (normalized) are what make it safe to plug into spectral GNN filters. The graph-select

Load-bearing premise

The operator keeps an edge whenever the tail potential x_u is at least the head potential x_v, applying this same orientation rule to positive and negative edges alike; the paper gives no justification or comparison for why negative edges should not follow the opposite rule, and if they should, the operator's quadratic form and the GNN built on it change.

What would settle it

Run NLSD-GNN on Bitcoin-OTC, Bitcoin-Alpha, Slashdot, and Epinions with the orientation rule for negative edges flipped (keep (u,v) when x_u <= x_v) while keeping everything else fixed; if the flipped variant matches or beats the paper's reported accuracies on SP/DP/3C/4C/5C, the central assumption that the same x_u >= x_v rule serves both signs is falsified. A smaller, directly checkable test: for the one-edge negative directed graph with x_u=1, x_v=-1 and normalized weights, the paper's quadratic form gives w(1-1)^2=0, while the flipped rule would give 4w; deciding which energy matches the i

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

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A single parameter-free operator now covers undirected, directed, signed, and signed-directed graphs; Table 3 frames the new Laplacian as the first in that list needing no extra parameters.
  • Because the normalized eigenvalues lie in [0,2], standard spectral GNN filters such as ChebNet-style polynomials apply without renormalization issues, which is the design used in NLSD-GNN.
  • The quadratic form gives a concrete optimization target: minimizing the energy pushes positive edges to be oriented along the potential order and negative edges to be cut, so the same operator can serve as an objective for clustering or ranking on signed-directed data.
  • The operator can be evaluated in O(M F d) time with sparse matrix multiplications, matching the complexity of linear spectral baselines while adding the graph-selection step.
  • On the four signed-directed link prediction benchmarks, the paper reports NLSD-GNN at or near the top for all five prediction tasks (SP, DP, 3C, 4C, 5C), including the combined sign-and-direction tasks.

Where Pith is reading between the lines

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

  • Beyond the paper: the choice to apply the same orientation rule x_u >= x_v to negative edges is an assumption; a natural variant flips the rule for negative edges (keep when x_u <= x_v), and comparing the two operators on the same benchmarks would isolate whether the sign-blind rule or a sign-aware rule better explains the gains.
  • Beyond the paper: because the potential vector x is produced by a trainable MLP, the learned potentials can be read as a node ranking; the operator could therefore be lifted directly to ranking-from-pairwise-comparisons problems, which the paper cites as motivation but does not benchmark.
  • Beyond the paper: the quadratic form suggests a signed-directed spectral-clustering objective with a Cheeger-type inequality; deriving such an inequality for L_G would extend the theory that exists for the unsigned directed nonlinear Laplacian.
  • Beyond the paper: the random-projection variant fixes the potential once at the start; an end-to-end version that recomputes the selected graph after each layer would test whether adaptivity of the subgraph selection is what drives the performance gain.

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

2 major / 7 minor

Summary. The paper defines a nonlinear Laplacian operator for signed-directed graphs, extending Yoshida's nonlinear digraph Laplacian. Given a potential vector x, the operator constructs a signed-undirected graph by keeping each directed edge (u,v) with the same sign when x_u >= x_v, and otherwise replacing the edge by same-sign self-loops of half weight; it then applies the standard signed-undirected Laplacian to the resulting graph. The paper proves that both the unnormalized and normalized operators are positive semidefinite and that the normalized operator has eigenvalues in [0,2]. It then proposes a spectral GNN, NLSD-GNN, with a fast random-projection variant and an MLP variant, and evaluates it on directed node classification and signed-directed link prediction. The theoretical results are clearly derived, and the experiments show strong performance on many benchmarks, though the empirical claim of universal superiority is not fully supported.

Significance. The operator fills a genuine gap: no nonlinear Laplacian for signed-directed graphs appears to have been studied, and the paper gives a clean extension of Yoshida's construction with the standard guarantees needed for spectral GNNs. The proofs are short, transparent, and correct modulo minor notational issues. The empirical study is broad, covering homophilic and heterophilic directed graphs, signed-directed link prediction, and several task formulations; if the design choices are robust, the method is a useful addition to the spectral GNN toolbox. The main weakness is that the orientation rule for negative edges is chosen without explicit justification or ablation, and the reported performance gains are often within one standard deviation, so the headline claim of 'superior performance' needs qualification.

major comments (2)
  1. [Definition 2 / Section 3.1] The rule x_u >= x_v is applied identically to positive and negative edges. For a negative directed edge, the equally natural alternative is to keep the edge when x_u <= x_v; this changes G_x and the quadratic form (the negative-edge term becomes 2x_u^2+2x_v^2 - max{x_v-x_u,0}^2). The same PSD and [0,2] eigenvalue arguments go through for the flipped variant. Since the message-passing graph and the quadratic form are different, every downstream NLSD-GNN result is contingent on this sign-blind orientation choice. The paper offers neither a semantic justification based on the meaning of signed direction nor an empirical ablation against the flipped rule. Please add a principled justification or a systematic comparison with the flipped-negative-edge variant, and adjust the empirical claims if the variants differ substantially.
  2. [Section 5.1 / Table 2] The abstract and Section 5.1 claim 'superior performance across diverse datasets.' Table 2 shows that NLSD-GNN-Fast is not strictly superior: it loses to MSGNN on Bitcoin-Alpha DP (70.5±0.5 vs 72.5±1.5) and ties on Bitcoin-OTC DP (71.8±0.7 vs 71.8±1.1). Moreover, many of the reported advantages are within one standard deviation of the best competitor (e.g., Bitcoin-Alpha 4C: 58.7±0.3 vs 58.5±0.7; Bitcoin-OTC 4C: 60.5±0.8 vs 59.8±0.7; Epinions DP: 86.7±0.6 vs 86.3±0.3). No paired significance tests or multiple-run comparisons are reported. The empirical claim should be softened to 'competitive or best on most tasks,' or significance testing should be added.
minor comments (7)
  1. [Appendix B, Theorem 1 proof] The line 'v^T(-lambda)v' should read 'v^T(lambda)v' (or simply 'lambda||v||^2'). The contradiction still works after fixing the sign.
  2. [Section 5.1 / Table 4] The text describes the datasets in Table 1 as 'homophilic,' but Table 4 lists Cornell, Texas, and Wisconsin with edge homophily 0.050. Please correct the categorization; these datasets are heterophilic.
  3. [Section 2] 'We discussed the related in the Appendix A' is an editing leftover and should be removed or rewritten.
  4. [Introduction] The sentence 'We begin by extending the definition of Yoshida [2016] ...' appears twice in Section 1. Please delete the duplicate.
  5. [Abstract] 'based on features More precisely' is missing punctuation; it should read 'based on features. More precisely, ...'
  6. [Appendix B / Notation] The expression 'x^T D^{-1/2} Q_G(x)(D^{-1/2} x)' is ambiguous; write it as 'Q_G(D^{-1/2} x)' or 'x^T D^{-1/2} L_G(D^{-1/2} x)'.
  7. [Section 3 / Normalized Laplacian] The proof of Theorem 2 divides by d_v. If the graph contains isolated vertices, a convention such as setting D^{-1/2}_{vv}=0 should be stated.

Circularity Check

0 steps flagged

No significant circularity; the main theoretical and empirical claims are self-contained, with only a non-load-bearing self-citation to Yoshida (2016).

full rationale

The paper's central theoretical claims (Definition 2 and Theorems 1 and 2) are derived directly from the definition of the nonlinear signed-directed Laplacian and from standard algebraic properties of signed/undirected Laplacians. The quadratic form expression for Q_G(x) is obtained by expanding x^T L_{G_x} x, and the positive-semidefiniteness and eigenvalue bounds follow from that expression; no fitted parameter, benchmark result, or external theorem is used as a substitute for the proof. The empirical claims are standard supervised benchmark evaluations, not predictions derived from fitted constants or from the theory itself. The only self-citation is to the same author's earlier nonlinear digraph Laplacian (Yoshida 2016), which motivates the construction and supplies the edge-dropout convention, but the present theorems do not reduce to that citation—they are proved from Definition 2. The sign-blind orientation rule x_u >= x_v for negative edges is an untested design choice that could affect empirical rankings, but that is a robustness/correctness concern, not circularity: flipping the rule would define a different operator, not the same one in disguise. No equation reduces to another by construction, and no fitted quantity is relabeled as a prediction. Therefore the paper is essentially self-contained with respect to circularity.

Axiom & Free-Parameter Ledger

2 free parameters · 4 axioms · 0 invented entities

The theoretical contribution is self-contained: the operator is defined, no free constants are fitted, and the proofs rest on standard signed-Laplacian facts. The empirical contribution depends on the unspecified random projection and MLP architecture choices, which are listed above. No new physical or mathematical entities are postulated.

free parameters (2)
  • Random projection vector r (Fast variant) = random, distribution unspecified
    Used in NLSD-GNN-Fast to compute node potentials x = H·r. The choice of r changes the edge dropout pattern and hence the results; the paper does not specify the distribution or seed.
  • MLP architecture for potential mapping (MLP variant) = not fully specified (halving dims until 1)
    The depth and hidden widths are not precisely given (Algorithm 2 halves dims but the number of layers depends on input dim); these are chosen by hand and affect results.
axioms (4)
  • standard math Signed Laplacian L^- = D^- + A^- is positive semidefinite.
    Used in Section 3 to define the signed Laplacian and in the proof of Theorem 1; established in Hou et al. 2003.
  • domain assumption The quadratic form of a self-loop in the signed Laplacian is 4w x_u^2 (equivalently, the self-loop contributes to L^- as D + A with diagonal entry w).
    Assumed implicitly in the expansion of Q_G for negative edges when x_u < x_v (Section 3.1); the paper does not state this convention explicitly.
  • domain assumption No parallel edges and nonnegative weights.
    Stated in Section 2.
  • standard math Eigenvalues of the nonlinear operator are characterized by L_G(v)=λv and v^T L_G(v)=λ||v||².
    Used in Appendix B for Theorem 2; this is the standard Rayleigh quotient argument applied to the nonlinear eigenpair definition.

pith-pipeline@v1.3.0-alltime-deepseek · 16596 in / 23639 out tokens · 244723 ms · 2026-08-05T00:12:52.193838+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of Nonlinear Laplacians Improve Signed-Directed Graph Learning." pith.science (2026). https://pith.science/paper/ZDIN7YUU

@misc{pith2026260800836,
  author       = {Pith},
  title        = {Pith review of: Nonlinear Laplacians Improve Signed-Directed Graph Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZDIN7YUU}},
  note         = {Machine review of arXiv:2608.00836}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

While signed-directed graphs have been studied using linear Laplacians in the design of graph neural networks, relatively little research has focused on developing non-linear Laplacian operators for such networks. We introduce a non-linear Laplacian operator specific to signed and directed networks (NLSD). This non-linear operator extends the concepts of the signed Laplacian for signed graphs and the Laplacian for directed graphs. The NLSD calculates node-specific potentials based on features More precisely, if the potential discrepancy is not aligned with the edge direction, we ignore it (and vice versa) leveraging message-passing techniques only across edges where potential discrepancies align with the edge's direction. Utilizing this novel operator, we propose an efficient spectral GNN framework (NLSD-GNN). We conducted comprehensive evaluations focusing on node classification and link prediction, examining scenarios involving signed, directional, or both types of information. Our findings reveal that this spectral GNN framework not only integrates signed and directional data effectively but also achieves superior performance across diverse datasets.

Figures

Figures reproduced from arXiv: 2608.00836 by Ali Parviz, Yuichi Yoshida.

Figure 1
Figure 1. Figure 1: The left figures show unsigned (1a) and signed (1b) directed graphs [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of our workflow, NLSD-GNN, with a toy example. In the first block, we use [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] 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

124 extracted references · 71 canonical work pages · 1 internal anchor

  1. [1]

    Hou, Yaoping and Li, Jiongsheng and Pan, Yongliang , journal=. On the. 2003 , publisher=

    2003

  2. [2]

    Computer Science

    Improving image generation with better captions , author=. Computer Science. https://cdn. openai. com/papers/dall-e-3. pdf , volume=

  3. [3]

    Neural Information Processing Systems , year=

    Interaction Networks for Learning about Objects, Relations and Physics , author=. Neural Information Processing Systems , year=

  4. [4]

    2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    Geometric Deep Learning on Graphs and Manifolds Using Mixture Model CNNs , author=. 2017 IEEE Conference on Computer Vision and Pattern Recognition (CVPR) , year=

    2017

  5. [5]

    Proceedings of the 2015 IEEE International Conference on Data Mining Workshop (ICDMW) , pages=

    Signed directed social network analysis applied to group conflict , author=. Proceedings of the 2015 IEEE International Conference on Data Mining Workshop (ICDMW) , pages=

    2015

  6. [6]

    Proceedings of the 52nd IEEE Conference on Decision and Control (CDC) , pages=

    Bipartite consensus for multi-agent systems on directed signed networks , author=. Proceedings of the 52nd IEEE Conference on Decision and Control (CDC) , pages=

  7. [7]

    Proceedings of the European Conference on Computer Vision (ECCV) , pages =

    Learning Human-Object Interactions by Graph Parsing Neural Networks , author=. Proceedings of the European Conference on Computer Vision (ECCV) , pages =

  8. [8]

    Generalized

    Maskey, Sohir and Parviz, Ali and Thiessen, Maximilian and St. Generalized. arXiv preprint arXiv:2210.15956 , year=

    Pith/arXiv arXiv

  9. [9]

    Neurally boosted supervised spectral clustering , author=

  10. [10]

    Proceedings of the 1st Workshop on NLP for Positive Impact , pages=

    Detecting hashtag hijacking for hashtag activism , author=. Proceedings of the 1st Workshop on NLP for Positive Impact , pages=

  11. [11]

    Learning on Graphs Conference , pages=

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

    2024

  12. [12]

    arXiv 2019 , author=

    Fake news detection on social media using geometric deep learning. arXiv 2019 , author=. arXiv preprint arXiv:1902.06673 , year=

    Pith/arXiv arXiv 2019

  13. [13]

    Msgnn: A spectral graph neural network based on a novel magnetic signed

    He, Yixuan and Perlmutter, Michael and Reinert, Gesine and Cucuringu, Mihai , booktitle=. Msgnn: A spectral graph neural network based on a novel magnetic signed

  14. [14]

    Information Sciences , volume=

    A new algorithm for positive influence maximization in signed networks , author=. Information Sciences , volume=. 2020 , publisher=

    2020

  15. [15]

    Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD) , pages=

    Graph convolutional neural networks for web-scale recommender systems , author=. Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining (KDD) , pages=

  16. [16]

    Bioinformatics , volume=

    Modeling polypharmacy side effects with graph convolutional networks , author=. Bioinformatics , volume=. 2018 , publisher=

    2018

  17. [17]

    Advances in Neural Information Processing Systems , volume=

    Link prediction based on graph neural networks , author=. Advances in Neural Information Processing Systems , volume=

  18. [18]

    Scientific Reports , volume=

    HyperFoods: Machine intelligent mapping of cancer-beating molecules in foods , author=. Scientific Reports , volume=. 2019 , publisher=

    2019

  19. [19]

    Nature Methods , volume=

    Deciphering interaction fingerprints from protein molecular surfaces using geometric deep learning , author=. Nature Methods , volume=. 2020 , publisher=

    2020

  20. [20]

    International Conference on Machine Learning (ICML) , pages=

    Gnnrank: Learning global rankings from pairwise comparisons via directed graph neural networks , author=. International Conference on Machine Learning (ICML) , pages=

  21. [21]

    Learning on Graphs Conference , pages=

    Pytorch geometric signed directed: A software package on graph neural networks for signed and directed graphs , author=. Learning on Graphs Conference , pages=

  22. [22]

    Detection and clustering of lead-lag networks for multivariate time series with an application to financial markets , author=

  23. [23]

    Sigmanet: One

    Fiorini, Stefano and Coniglio, Stefano and Ciavotta, Michele and Messina, Enza , booktitle=. Sigmanet: One

  24. [24]

    Proceedings of the 28th International Conference on Artificial Neural Networks (ICANN) , pages=

    Signed graph attention networks , author=. Proceedings of the 28th International Conference on Artificial Neural Networks (ICANN) , pages=

  25. [25]

    Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume=

    Learning signed network embedding via graph attention , author=. Proceedings of the AAAI Conference on Artificial Intelligence (AAAI) , volume=

  26. [26]

    Signed Graph Neural Networks: A Frequency Perspective

    Signed graph neural networks: A frequency perspective , author=. arXiv preprint arXiv:2208.07323 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  27. [27]

    arXiv preprint arXiv:1911.03082 , year=

    Composition-based multi-relational graph convolutional networks , author=. arXiv preprint arXiv:1911.03082 , year=

    Pith/arXiv arXiv 1911

  28. [28]

    Proceedings of the SIGCHI conference on human factors in computing systems , pages=

    Signed networks in social media , author=. Proceedings of the SIGCHI conference on human factors in computing systems , pages=

  29. [29]

    Proceedings of the AAAI conference on Artificial Intelligence (AAAI) , volume=

    Relational graph neural network with hierarchical attention for knowledge graph completion , author=. Proceedings of the AAAI conference on Artificial Intelligence (AAAI) , volume=

  30. [30]

    Proceedings of the 15th Extended Semantic Web Conference (ESWC) , pages=

    Modeling relational data with graph convolutional networks , author=. Proceedings of the 15th Extended Semantic Web Conference (ESWC) , pages=

  31. [31]

    , author=

    On the notion of balance of a signed graph. , author=. Michigan Mathematical Journal , volume=. 1953 , publisher=

    1953

  32. [32]

    Proceedings of the 2018 IEEE International Conference on Data Mining (ICDM) , pages=

    Signed graph convolutional networks , author=. Proceedings of the 2018 IEEE International Conference on Data Mining (ICDM) , pages=

    2018

  33. [33]

    Proceedings of the 2022 SIAM International Conference on Data Mining (SDM) , pages=

    SSSNET: semi-supervised signed network clustering , author=. Proceedings of the 2022 SIAM International Conference on Data Mining (SDM) , pages=

    2022

  34. [34]

    Proceedings of the AAAI conference on Artificial Intelligence (AAAI) , volume=

    SDGNN: Learning node representation for signed directed networks , author=. Proceedings of the AAAI conference on Artificial Intelligence (AAAI) , volume=

  35. [35]

    ArXiv , year=

    Graph Attention Networks , author=. ArXiv , year=

  36. [36]

    ArXiv , year=

    Convolutional Networks on Graphs for Learning Molecular Fingerprints , author=. ArXiv , year=

  37. [37]

    Neural Information Processing Systems , year=

    Inductive Representation Learning on Large Graphs , author=. Neural Information Processing Systems , year=

  38. [38]

    Neural Information Processing Systems , year=

    Diffusion-Convolutional Neural Networks , author=. Neural Information Processing Systems , year=

  39. [39]

    Neural Computation , year=

    Laplacian Eigenmaps for Dimensionality Reduction and Data Representation , author=. Neural Computation , year=

  40. [40]

    ArXiv , year=

    PyTorch Geometric Signed Directed: A Survey and Software on Graph Neural Networks for Signed and Directed Graphs , author=. ArXiv , year=

  41. [41]

    International Conference on Machine Learning (ICML) , year=

    Neural Message Passing for Quantum Chemistry , author=. International Conference on Machine Learning (ICML) , year=

  42. [42]

    IEEE Transactions on Neural Networks , year=

    The Graph Neural Network Model , author=. IEEE Transactions on Neural Networks , year=

  43. [43]

    International Conference on Learning Representations (ICLR) , year=

    Gated Graph Sequence Neural Networks , author=. International Conference on Learning Representations (ICLR) , year=

  44. [44]

    Advances in Neural Information Processing Systems , volume=

    Diffusion-Convolutional Neural Networks , author=. Advances in Neural Information Processing Systems , volume=

  45. [45]

    Neural computation , volume =

    Belkin, Mikhail and Niyogi, Partha , title =. Neural computation , volume =

  46. [46]

    Science , volume =

    Benson, Austin R and Gleich, David F and Leskovec, Jure , title =. Science , volume =

  47. [47]

    ICLR Workshop on Representation Learning on Graphs and Manifolds , year =

    Bojchevski, Aleksandar and Günnemann, Stephan , title =. ICLR Workshop on Representation Learning on Graphs and Manifolds , year =

  48. [48]

    Bovet, Alexandre and Grindrod, Peter , title =

  49. [49]

    International Conference on Learning Representations (ICLR) , year =

    Bruna, Joan and Zaremba, Wojciech and Szlam, Arthur and LeCun, Yann , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  50. [50]

    Annals of Combinatorics , volume =

    Chung, Fan , title =. Annals of Combinatorics , volume =

    work page

  51. [51]

    International Workshop on Algorithms and Models for the Web-Graph , pages =

    Chung, Fan and Kempton, Mark , title =. International Workshop on Algorithms and Models for the Web-Graph , pages =

    work page

  52. [52]

    Chung, Fan RK and Chung Graham, Fan , title =

    work page

  53. [53]

    Applied and Computational Harmonic Analysis , volume =

    Cloninger, Alexander , title =. Applied and Computational Harmonic Analysis , volume =

    work page

  54. [54]

    Nonlinear

    Yuichi Yoshida , journal=. Nonlinear

    work page

  55. [55]

    Applied and computational harmonic analysis , volume =

    Coifman, Ronald R and Lafon, Stéphane , title =. Applied and computational harmonic analysis , volume =

    work page

  56. [56]

    International Conference on Artificial Intelligence and Statistics , pages =

    Cucuringu, Mihai and Li, Huan and Sun, He and Zanetti, Luca , title =. International Conference on Artificial Intelligence and Statistics , pages =

    work page

  57. [57]

    Advances in Neural Information Processing Systems , pages =

    Defferrard, Michaël and Bresson, Xavier and Vandergheynst, Pierre , title =. Advances in Neural Information Processing Systems , pages =

    work page

  58. [58]

    Advances in Neural Information Processing Systems , volume =

    Duvenaud, David K and Maclaurin, Dougal and Iparraguirre, Jorge and Bombarell, Rafael and Hirzel, Timothy and Aspuru-Guzik, Alan and Adams, Ryan P , title =. Advances in Neural Information Processing Systems , volume =

    work page

  59. [59]

    and da F

    de Resende, Bruno Messias F. and da F. Costa, Luciano , title =

    work page

  60. [60]

    Physical Review E , volume =

    Fanuel, Michaël and Alaiz, Carlos M and Suykens, Johan AK , title =. Physical Review E , volume =

    work page

  61. [61]

    ArXiv , year=

    Magnetic Eigenmaps for Visualization of Directed Networks , author=. ArXiv , year=

    work page

  62. [62]

    Machine Learning and Knowledge Discovery in Databases , pages =

    Furutani, Satoshi and Shibahara, Toshiki and Akiyama, Mitsuaki and Hato, Kunio and Aida, Masaki , title =. Machine Learning and Knowledge Discovery in Databases , pages =

    work page

  63. [63]

    Journal of Graph Theory , volume =

    Guo, Krystal and Mohar, Bojan , title =. Journal of Graph Theory , volume =

    work page

  64. [64]

    Advances in Neural Information Processing Systems , year=

    MagNet: A Neural Network for Directed Graphs , author=. Advances in Neural Information Processing Systems , year=

    work page

  65. [65]

    MSGNN: A Spectral Graph Neural Network Based on a Novel Magnetic Signed

    Yixuan He and Michael Permultter and Gesine Reinert and Mihai Cucuringu , booktitle=. MSGNN: A Spectral Graph Neural Network Based on a Novel Magnetic Signed

    work page

  66. [66]

    and Ying, Rex and Leskovec, Jure , title =

    Hamilton, William L. and Ying, Rex and Leskovec, Jure , title =. Proceedings of the 31st International Conference on Neural Information Processing Systems , pages =. 2017 , address =

    work page 2017

  67. [67]

    Applied and Computational Harmonic Analysis , volume =

    Hammond, David K and Vandergheynst, Pierre and Gribonval, Rémi , title =. Applied and Computational Harmonic Analysis , volume =

    work page

  68. [68]

    ArXiv , year=

    Wavelets on Graphs via Spectral Graph Theory , author=. ArXiv , year=

    work page

  69. [69]

    Naganand Yadati and Madhav Nimishakavi and Prateek Yadav and Vikram Nitin and Anand Louis and Partha Pratim Talukdar , journal=

    work page

  70. [70]

    Proceedings of the 47th Annual ACM symposium on Theory of Computing , pages=

    Hypergraph Markov Operators, Eigenvalues and Approximation Algorithms , author=. Proceedings of the 47th Annual ACM symposium on Theory of Computing , pages=

    work page

  71. [71]

    and Welling, Max , title =

    Kipf, Thomas N. and Welling, Max , title =. International Conference on Learning Representations (ICLR) , year =

  72. [72]

    International Conference on Learning Representations (ICLR) , year =

    Klicpera, Johannes and Bojchevski, Aleksandar and Günnemann, Stephan , title =. International Conference on Learning Representations (ICLR) , year =

    work page

  73. [73]

    Levie, Ron and Huang, Wei and Bucci, Lorenzo and Bronstein, Michael M and Kutyniok, Gitta , title =

    work page

  74. [74]

    Statistical Mechanics , pages =

    Lieb, Elliott H and Loss, Michael , title =. Statistical Mechanics , pages =

    work page

  75. [75]

    Ma, Yi and Hao, Jianye and Yang, Yaodong and Li, Han and Jin, Junqi and Chen, Guangyong , title =

    work page

  76. [76]

    IEEE Signal Processing Magazine , volume =

    Marques, Antonio G and Segarra, Santiago and Mateos, Gonzalo , title =. IEEE Signal Processing Magazine , volume =

    work page

  77. [77]

    Mernyei, Péter and Cangea, Catălina , title =

    work page

  78. [78]

    Linear Algebra and its Applications , volume =

    Mohar, Bojan , title =. Linear Algebra and its Applications , volume =

    work page

  79. [79]

    , title =

    Monti, Federico and Otness, Karl and Bronstein, Michael M. , title =. 2018 IEEE Data Science Workshop , pages =

    work page 2018

  80. [80]

    Proceedings of the IEEE , volume =

    Ortega, Antonio and Frossard, Pascal and Kovacevic, Jelena and Moura, José MF and Vandergheynst, Pierre , title =. Proceedings of the IEEE , volume =

    work page

Showing first 80 references.