Pith. sign in

REVIEW 3 major objections 5 minor 60 references

GegenNet claims a Gegenbauer-polynomial spectral filter outperforms monomial filters for link sign prediction in signed bipartite graphs.

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 ↗

GegenNet predicts link signs in signed bipartite graphs with Gegenbauer-polynomial spectral filters on positive and negative edges, reporting up to 4.28% AUC and 11.69% F1 gains over 11 baselines.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection GegenNet's Gegenbauer filter idea is worth a look, but the empirical claims are compromised by test-set leakage in filter selection and a wrong equivalence in the initialization derivation. the 3 major comments →

arxiv 2508.19907 v1 pith:6BX2C4W2 submitted 2025-08-27 cs.LG cs.SI

GegenNet: Spectral Convolutional Neural Networks for Link Sign Prediction in Signed Bipartite Graphs

Hewen Wang , Renchi Yang , Xiaokui Xiao This is my paper
classification cs.LG cs.SI MSC 68T0705C8268R10
keywords link sign predictionsigned bipartite graphsspectral graph filtersGegenbauer polynomialsgraph neural networksnode embedding initializationspectral decompositionsigned graph representation learning
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 reading

The paper tries to show that link sign prediction in signed bipartite graphs is best done with a spectral convolutional network whose filters are built from Gegenbauer polynomials rather than the monomial or Chebyshev filters used by prior graph neural networks. It recasts the prediction task as four curve-fitting problems mapping the spectra of the positive and negative adjacency matrices to the missing positive and negative links, and reports that on real datasets the ideal fitting curves bend sharply at low and high frequencies, exactly where monomial filters (GCN, APPNP, GDC) deviate. GegenNet’s contribution is a filter family that fits those bends, applied through separate positive and negative convolutional streams plus a spectral feature initialization. On six benchmark SBG datasets the model reports the highest AUC and F1 against 11 baselines, with gains up to 4.28% AUC and 11.69% F1. If these results reproduce, the paper’s core claim is that orthogonal polynomial spectral filters are a better foundation for signed bipartite link prediction than the standard monomial ones.

Core claim

The central claim is that a spectral convolutional network using Gegenbauer-polynomial filters, applied separately to the positive and negative adjacency matrices, captures low- and high-frequency spectral signals that monomial-based filters miss, and that this leads to state-of-the-art link sign prediction on signed bipartite graphs. The paper derives the task as four spectral curve-fitting problems (positive-to-positive, negative-to-positive, negative-to-negative, positive-to-negative), observes that ideal target curves bend sharply outside the mid-frequency range, and shows that the Gegenbauer basis, generalizing Chebyshev and Legendre polynomials, fits those curves better than the spectr

What carries the argument

The load-bearing object is the Gegenbauer polynomial spectral filter J_k^alpha(A), defined by the recurrence J_0 = I, J_1 = (alpha+1)A, and J_k = omega_k A J_{k-1} - omega'_k J_{k-2} for k>=2, with coefficients omega_k and omega'_k depending on alpha and k. This filter takes the place of the monomial basis in the four spectral curve-fitting functions, letting the network fit low-, mid-, and high-frequency graph signals that monomial bases distort. The convolutional layer produces three parallel embeddings H_pos, H_neg, and H_org by applying the Gegenbauer filter to A+, to A-, and to the identity, then fuses them by concatenation and a linear map; the paper shows that after linearization, the

Load-bearing premise

The load-bearing premise is that the inter-partition objective in Eq. (12), which explicitly pulls positive edges together and pushes negative edges apart, is equivalent to the trace objective with L = D - (A+ + A-), the Laplacian of the unsigned adjacency. Since that Laplacian assigns weight 1 to both positive and negative edges, the equivalence discards the sign information it claims to encode, so the eigenvector initialization computed from it is not sign-aware as stated.

What would settle it

Compute the Eq. (12) objective and trace(X^T L X) with L = D - (A+ + A-) on any SBG with at least one negative edge; the two values differ whenever negative edges exist, so a direct numerical check settles whether the initialization is actually sign-aware. A second experiment replaces the Laplacian in the feature initialization with the signed Laplacian D - (A+ - A-) and compares AUC/F1: if results change materially, the reported gains are not attributable to the claimed sign-aware spectral features.

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

If this is right

  • If GegenNet reproduces, Gegenbauer-polynomial spectral filters become a strong default choice for signed bipartite link prediction, outperforming monomial, Chebyshev, and Legendre bases on the tested benchmarks.
  • The four-curve-fitting view implies that link sign prediction can be decomposed into separate spectral transfer functions for positive-to-positive, negative-to-positive, negative-to-negative, and positive-to-negative inference, each of which is better approximated by an orthogonal polynomial basis.
  • The ablation results indicate that the positive and negative spectral streams are both necessary, with the negative stream playing a particularly large role on datasets with few negative edges, such as Bonanza.
  • Spectral feature initialization from the Laplacian eigenvectors and the singular vectors of the normalized bi-adjacency matrix contributes substantially more than random initialization, suggesting that topology-first initialization helps in signed bipartite settings.
  • The linearization analysis shows that depth L yields up to 3^L distinct filter combinations, so the model’s capacity grows combinatorially with layers rather than merely linearly.

Where Pith is reading between the lines

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

  • The paper leaves implicit that alpha, the Gegenbauer parameter, is a continuous spectrum between Chebyshev (alpha=0), Legendre (alpha=0.5), and higher-order variants, so dataset-specific alpha search could further improve results on imbalanced datasets.
  • The sign-awareness of the feature initialization depends on the equivalence between the signed inter-partition objective and the trace objective with an unsigned Laplacian; a testable variant replaces that Laplacian with a signed Laplacian D-(A+ - A-) and checks whether the reported gains persist, an experiment the paper does not run.
  • The four curve-fitting formulations suggest a natural extension to directed signed edges by orienting the four transfer functions separately for outgoing and incoming relations, which the paper does not explore despite its three-stream architecture.
  • The 3^L decomposition of the linearized model hints that for very large graphs, approximating the Gegenbauer polynomial products with random features or a fixed sparsification could scale the method beyond the tested datasets.
Share X Bluesky LinkedIn Reddit HN

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 / 5 minor

Summary. The paper proposes GegenNet, a spectral graph neural network for link sign prediction in signed bipartite graphs (SBGs). The method has three claimed contributions: (i) a spectral feature initialization scheme that combines eigenvectors of the graph Laplacian with left singular vectors of a normalized bi-adjacency matrix; (ii) a spectral graph filter based on Gegenbauer polynomials, motivated by a curve-fitting analysis of ideal spectral filters on two datasets; and (iii) multi-layer sign-aware convolutional layers that apply separate Gegenbauer filters to positive and negative adjacency matrices and fuse them. Experiments compare GegenNet against 11 baselines on 6 SBG datasets and report consistent improvements, up to +4.28% AUC and +11.69% F1. The paper also includes ablation studies and hyperparameter sensitivity analyses.

Significance. If the reported results are valid, GegenNet would be a strong new baseline for link sign prediction on signed bipartite graphs, and the Gegenbauer-polynomial filter would be a useful addition to the spectral GNN toolbox. The curve-fitting reformulation in Lemma 4.1 is a nice idea, and the public code/datasets are a strength. However, the empirical claim is currently undermined by a test-set leakage concern in the filter design and hyperparameter selection, and the spectral initialization derivation contains a concrete mathematical error. These issues are load-bearing for the paper's central claims, so the significance can only be assessed after a careful revision.

major comments (3)
  1. [§4.1.1, §5.1, Figs. 1–4, Table 4] The design of the Gegenbauer filter is informed by held-out labels. In §4.1.1, Y+ and Y− are introduced as the missing links, and Lemma 4.1 together with Fig. 1 fit the 'ideal' spectral curves U+ᵀY+U+, U+ᵀY−U+, etc. on MLM and House1to10. Under the 8:1:1 train/validation/test split described in §5.1, these missing links are the test edges of the same datasets that appear in Table 4. The conclusion that monomial filters are inadequate and that the Gegenbauer basis fits best is therefore drawn from the test labels. Figs. 2–4 also report AUC versus μ, L, and d without stating that those curves are computed on a validation split. If α, μ, L, and d were selected using test AUC, the gains in Table 4 are not a valid estimate of generalization. Please re-run the model selection using only training/validation data (or nested cross-validation) and report test results for the model chosen without t
  2. [§4.3, Eqs. (12)–(13)] The claimed equivalence between Eq. (12) and Eq. (13) is incorrect. Eq. (12) minimizes distances on positive edges and maximizes distances on negative edges, i.e., it assigns weight +1 to E+ and −1 to E−. Eq. (13), however, uses ilde A = [0 A; Aᵀ 0] with A = A+ + A−, so every known edge contributes +1 to the sum and the sign information is discarded. The unsigned Laplacian trace objective is not equal to the signed objective. Consequently, the optimal X given by the d smallest eigenvectors of the unsigned Laplacian L does not follow from Eq. (12), and the claim that X = μΦ + (1−μ)Ψ is a 'sign-aware spectral initialization' is not supported. A signed Laplacian (e.g., D_s − (A+ − A−) in appropriate block form) would be required, which would change the proposed features.
  3. [§5.2, Table 4, §4.1.3, Table 2] There is an internal inconsistency in the choice of α and in the claim that α=1.5 is consistently superior. Section 4.1.3 states that Gegenbauer with α=1.0 and k=3 fits the ideal curves best, while Table 2 lists α=1.5 for 'Ours'. Section 5.2 claims that 'the Gegenbauer-based filtering with α=1.5 consistently demonstrates superior adaptability', but Table 4 shows GegenNet(α=0) has higher or equal AUC than GegenNet(α=1.5) on Amazon (0.8017 vs 0.8012), House1to10 (0.9269 vs 0.9260), and Review (0.7952 vs 0.7857), and Table 5 shows similar patterns. The choice of α appears to be an artifact of test-set tuning rather than a principled, pre-specified setting. This needs to be resolved and reported transparently.
minor comments (5)
  1. [§5.2, Table 4] Several strong baselines (SBGCL, SBGNN, SLGNN) are marked '-' on Amazon and MLM with the explanation that they 'encounter errors during execution or fail to converge'. This makes the head-to-head comparison incomplete. Please clarify whether these failures are due to memory, implementation, or hyperparameter issues, and consider reporting results where available.
  2. [§4.1.3, Theorem 4.2] The statement 'For K ≥ N0' should presumably be 'For K ∈ N0'. Also, the proof of Theorem 4.2 relies on Jacobi polynomial identities but does not explicitly state the required parameter ranges; please make the conditions precise.
  3. [§3.2, Table 1] There are minor typos: 'Proportition 3.12' should be 'Proposition 3.1', and 'Symbol and and Terminology' in §3.1 should be 'Symbols and Terminology'.
  4. [§5.4, Figs. 2–4] The hyperparameter plots would be more informative if they indicated the number of runs / standard deviations, and if the axis labels for d in Fig. 4 were consistent (the text mentions truncation to 128 for Review and Senate, but the plots appear to start at d=22 or 23 without explaining the truncated cases).
  5. [§4.2, Linearization Analysis] The linearization analysis removes the PReLU nonlinearity, but the actual model uses PReLU. The decomposition into 3^K terms is therefore only an approximation of the true forward pass. Please state this explicitly when using the linearization to argue model capacity.

Circularity Check

1 steps flagged

Test-set-guided design of the Gegenbauer filter makes the SOTA claim substantially circular; Sec. 4.3 Eq. (12)/(13) is a separate sign-information bug.

specific steps
  1. fitted input called prediction [Sec. 4.1.1–4.1.3 (Eqs. 4–6, Fig. 1, Table 2); evaluation in Sec. 5.2 (Table 4)]
    "Let Y+ and Y− contain the missing positive and negative links in the existing adjacency matrices Â+ and Â−, respectively. Inspired by this, on the MLM and House1to10 datasets, we empirically study the choices of f(++)(·), f(−−)(·), f(−+)(·), and f(+−)(·) ... As plotted in Fig. 1, we can observe that the Gegenbauer polynomial basis Jαk(·) well fits the target curves in the four cases when α = 1.0 and k = 3, demonstrating its high efficacy in link sign prediction on SBGs."

    Y+ and Y− are the missing/test links that the model is supposed to predict. The paper uses U⊤Y U curves derived from these labels on MLM and House1to10 to compare existing spectral filters and to conclude that the Gegenbauer basis (with α and k) 'well fits' the ideal curves, i.e., the model family and hyperparameters are selected to fit the test labels. The same two datasets appear in Table 4 under the 8:1:1 train/validation/test protocol of Sec. 5.1, where GegenNet's reported gains over competitors are presented as predictions. Thus the central design choice is fitted to the evaluation target; the headline SOTA numbers are not an out-of-sample test of that choice. The inconsistency between α=1.0 in Sec. 4.1.3 and α=1.5 in Table 2 confirms the α choice was not fixed a priori.

full rationale

The principal circular element is the use of test labels to motivate and select the Gegenbauer spectral filter family before reporting results on those same datasets. That is a fitted-input-called-prediction pattern and warrants a score of 6: the central model-family design is fitted to the evaluation target, so the reported SOTA claim is substantially compromised, although the trained weights and the α=0/0.5 variants still provide partial independent evidence. I find no load-bearing self-citation: the authors' prior work (EAGLE [49], etc.) is cited for background, not as the justification for the Gegenbauer filter or for a uniqueness claim. Theorem 4.2 rests on an external Jacobi-polynomial result [44], and Lemma 4.1 on an external spectral transformation result [30], so those are not circular. Separately, Sec. 4.3 asserts that Eq. (12), which minimizes positive-edge distances and maximizes negative-edge distances, is equivalent to trace(X^T L X) with L = D − \tilde A and \tilde A = A+ + A−; since \tilde A assigns weight +1 to both positive and negative edges, the sign information in Eq. (12) is discarded, so the claimed equivalence is false. That is a correctness/validity bug, not a circularity, and I do not count it in the score, but it further weakens the 'sign-aware spectral initialization' contribution. Figs. 2–4 likewise do not state that AUC curves are computed on the validation split, so the μ/L/d selection may be an additional source of test leakage, reinforcing the main step.

Axiom & Free-Parameter Ledger

5 free parameters · 5 axioms · 0 invented entities

The central model relies on standard spectral graph theory, but the paper introduces a mathematically incorrect bridge between its signed objective and the unsigned Laplacian, and the choice of the Gegenbauer basis is empirically fitted to test-derived ideal curves. There are no new physical entities.

free parameters (5)
  • alpha (Gegenbauer parameter) = 0, 0.5, 1.5 (dataset-dependent best; Fig. 1 uses 1.0)
    Selects the polynomial family; chosen by hand after inspecting spectral fit curves computed from test labels.
  • mu (feature combination weight) = 0.3
    Weights the inter- vs intra-partition spectral features in Eq. (17); tuned as a hyperparameter (Fig. 2).
  • L (number of layers) = 3
    Depth of the sign-aware convolutional stack; tuned in Fig. 3.
  • d (SVD/eigenvector dimension) = 32 (truncated to 128 on Review/Senate)
    Dimension of spectral node features; tuned in Fig. 4.
  • delta (layer scaling) = not specified
    Weight hyperparameter in Eq. (9); no value given in Section 5.1.
axioms (5)
  • standard math Ky Fan trace minimization (Thm 4.3) for deriving the inter-partition feature basis
    Unproved background from [7]; used to justify choosing the d smallest eigenvectors of L.
  • ad hoc to paper The claimed equivalence in Eq. (12) to Eq. (13): signed distance objective equals unsigned Laplacian trace
    This is incorrect; the objective in Eq. (12) subtracts negative-edge distances, but \tilde A has entries 1 for both signs, so the trace uses positive weights only. The initialization is consequently not sign-aware.
  • domain assumption The four-function reformulation of link sign prediction (Eqs. 4-5)
    Assumes the sign prediction task can be reduced to four independent least-squares fittings from known positive/negative edges to missing positive/negative signs; reasonable but a modeling choice.
  • standard math The spectral filter heuristics (CN, k-hop RW, PPR, HKPR) unify as f(Lambda) (Prop 3.1)
    Standard spectral decomposition of polynomial/geometric/heat-kernel functions of the adjacency matrix.
  • ad hoc to paper Gegenbauer polynomials fit the observed ideal spectral curves better than monomial filters
    Based on visual inspection of Fig. 1 on two datasets, with ideal curves computed from held-out labels; not proven and partly selected on test data.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of GegenNet: Spectral Convolutional Neural Networks for Link Sign Prediction in Signed Bipartite Graphs." pith.science (2026). https://pith.science/paper/6BX2C4W2

@misc{pith2026250819907,
  author       = {Pith},
  title        = {Pith review of: GegenNet: Spectral Convolutional Neural Networks for Link Sign Prediction in Signed Bipartite Graphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6BX2C4W2}},
  note         = {Machine review of arXiv:2508.19907}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Given a signed bipartite graph (SBG) G with two disjoint node sets U and V, the goal of link sign prediction is to predict the signs of potential links connecting U and V based on known positive and negative edges in G. The majority of existing solutions towards link sign prediction mainly focus on unipartite signed graphs, which are sub-optimal due to the neglect of node heterogeneity and unique bipartite characteristics of SBGs. To this end, recent studies adapt graph neural networks to SBGs by introducing message-passing schemes for both inter-partition (UxV) and intra-partition (UxU or VxV) node pairs. However, the fundamental spectral convolutional operators were originally designed for positive links in unsigned graphs, and thus, are not optimal for inferring missing positive or negative links from known ones in SBGs. Motivated by this, this paper proposes GegenNet, a novel and effective spectral convolutional neural network model for link sign prediction in SBGs. In particular, GegenNet achieves enhanced model capacity and high predictive accuracy through three main technical contributions: (i) fast and theoretically grounded spectral decomposition techniques for node feature initialization; (ii) a new spectral graph filter based on the Gegenbauer polynomial basis; and (iii) multi-layer sign-aware spectral convolutional networks alternating Gegenbauer polynomial filters with positive and negative edges. Our extensive empirical studies reveal that GegenNet can achieve significantly superior performance (up to a gain of 4.28% in AUC and 11.69% in F1) in link sign prediction compared to 11 strong competitors over 6 benchmark SBG datasets.

Figures

Figures reproduced from arXiv: 2508.19907 by Hewen Wang, Renchi Yang, Xiaokui Xiao.

Figure 1
Figure 1. Figure 1: The relations between Λ + and U +⊤Y +U + , 𝚲 − and U −⊤Y −U −, 𝚲 + and U +⊤Y −U + , 𝚲 − and U −⊤Y +U − [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: AUC by varying 𝜇 in GegenNet. GegenNet (𝛼 = 0, Chebyshev) GegenNet (𝛼 = 1.5) GegenNet (𝛼 = 0.5, Legendre) 2 3 4 5 6 7 0.77 0.78 0.79 0.80 AUC (a) Amazon 2 3 4 5 6 7 0.66 0.68 0.70 0.72 AUC (b) Bonanza 2 3 4 5 6 7 0.918 0.920 0.922 0.924 0.926 AUC (c) House1to10 2 3 4 5 6 7 0.802 0.804 0.806 0.808 0.810 AUC (d) MLM 2 3 4 5 6 7 0.71 0.72 0.73 0.74 AUC (e) Review 2 3 4 5 6 7 0.895 0.900 0.905 AUC (f) Senate … view at source ↗
Figure 3
Figure 3. Figure 3: AUC by varying 𝐿 in GegenNet. GegenNet (𝛼 = 0, Chebyshev) GegenNet (𝛼 = 1.5) GegenNet (𝛼 = 0.5, Legendre) 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 0.78 0.79 0.80 0.81 AUC (a) Amazon 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 0.66 0.68 0.70 0.72 AUC (b) Bonanza 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 0.89 0.90 0.91 0.92 AUC (c) House1to10 2 2 2 3 2 4 2 5 2 6 2 7 2 8 2 9 0.800 0.802 0.804 0.806 0.808 0.810 AUC (d) MLM 2 2 2 3 2 4 2 5 2… view at source ↗
Figure 4
Figure 4. Figure 4: AUC by varying 𝑑 in GegenNet. Furthermore, GegenNet shows strong resilience on sparse and imbalanced datasets. For example, the Bonanza dataset contains very few negative links, while the Review dataset is characterized by sparsely connected nodes. Despite these challenges, GegenNet maintains robust performance, achieving competitive AUC and F1 scores where other methods tend to struggle. Overall, GegenNet… 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

60 extracted references · 51 canonical work pages · 1 internal anchor

  1. [1]

    Jiangxia Cao, Xixun Lin, Shu Guo, Luchen Liu, Tingwen Liu, and Bin Wang. 2021. Bipartite Graph Embedding via Mutual Information Maximization. In WSDM. New York, NY, USA, 635–643. https://doi.org/10.1145/3437963.3441783

    arXiv 2021

  2. [2]

    Dorwin Cartwright and Frank Harary. 1956. Structural balance: a generalization of Heider’s theory. Psychological review 63, 5 (1956), 277

    work page 1956

  3. [3]

    Fan Chung. 2007. The heat kernel as the pagerank of a graph. Proceedings of the National Academy of Sciences 104, 50 (2007), 19735–19740

    work page 2007

  4. [4]

    Mihai Cucuringu, Peter Davies, Aldo Glielmo, and Hemant Tyagi. 2020. Sponge: A generalized eigenproblem for clustering signed networks. In AISTATS 2019 - 22nd Int. Conf. Artif. Intell. Stat. (Proceedings of Machine Learning Research, Vol. 89), Kamalika Chaudhuri and Masashi Sugiyama (Eds.). PMLR, 1088–1098. arXiv:1904.08575 https://proceedings.mlr.press/v...

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  5. [5]

    Michaël Defferrard, Xavier Bresson, and Pierre Vandergheynst. 2016. Convolu- tional neural networks on graphs with fast localized spectral filtering. Advances in neural information processing systems 29 (2016)

    2016

  6. [6]

    Tyler Derr, Yao Ma, and Jiliang Tang. 2018. Signed graph convolutional networks. In 2018 IEEE International Conference on Data Mining (ICDM) . IEEE, 929–934

    work page 2018

  7. [7]

    Ky Fan. 1949. On a theorem of Weyl concerning eigenvalues of linear transfor- mations I. Proceedings of the National Academy of Sciences of the United States of America 35, 11 (1949), 652

    work page 1949

  8. [8]

    Stefano Fiorini, Stefano Coniglio, Michele Ciavotta, and Enza Messina. 2023. SigMaNet: one laplacian to rule them all. In AAAI. 9 pages

    work page 2023

  9. [9]

    Ming Gao, Leihui Chen, Xiangnan He, and Aoying Zhou. 2018. BiNE: Bipartite network embedding. In SIGIR. 715–724

    work page 2018

  10. [10]

    Johannes Gasteiger, Stefan Weißenberger, and Stephan Günnemann. 2019. Diffu- sion improves graph learning. NeurIPS 32 (2019)

    work page 2019

  11. [11]

    Mark S Granovetter. 1973. The strength of weak ties.American journal of sociology 78, 6 (1973), 1360–1380

    work page 1973

  12. [12]

    Chaoyang He, Tian Xie, Yu Rong, Wenbing Huang, Junzhou Huang, Xiang Ren, Cyrus Shahabi, and Xi-Ang Ren. 2020. Cascade-BGNN: Toward Efficient Self- supervised Representation Learning on Large-scale Bipartite Graphs . Technical Report. https://doi.org/10.1145/1122445.1122456 arXiv:1906.11994v3

    work page arXiv 2020

  13. [13]

    Yixuan He, Michael Perlmutter, Gesine Reinert, and Mihai Cucuringu. 2022. Msgnn: A spectral graph neural network based on a novel magnetic signed laplacian. In Learning on Graphs Conference . PMLR, 40–1

    work page 2022

  14. [14]

    2022.SSSNET: Semi-Supervised Signed Network Clustering

    Yixuan He, Gesine Reinert, Songchao Wang, and Mihai Cucuringu. 2022.SSSNET: Semi-Supervised Signed Network Clustering . 244–252

    work page 2022

  15. [15]

    Yixuan He, Gesine Reinert, Songchao Wang, and Mihai Cucuringu. 2022. SSSNET: Semi-Supervised Signed Network Clustering. In SDM. 244–252

    work page 2022

  16. [16]

    Junjie Huang, Huawei Shen, Qi Cao, Shuchang Tao, and Xueqi Cheng. 2021. Signed bipartite graph neural networks. In Proceedings of the 30th ACM Interna- tional Conference on Information & Knowledge Management . 740–749

    work page 2021

  17. [17]

    Junjie Huang, Huawei Shen, Liang Hou, and Xueqi Cheng. 2019. Signed Graph Attention Networks. Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) 11731 LNCS (jun 2019), 566–577. https://doi. org/10.1007/978-3-030-30493-5_53 arXiv:1906.10958

    Pith/arXiv arXiv 2019

  18. [18]

    Junjie Huang, Huawei Shen, Liang Hou, and Xueqi Cheng. 2019. Signed graph attention networks. In Artificial Neural Networks and Machine Learning–ICANN 2019: Workshop and Special Sessions . 566–577

    work page 2019

  19. [19]

    Junjie Huang, Huawei Shen, Liang Hou, and Xueqi Cheng. 2021. SDGNN: Learn- ing Node Representation for Signed Directed Networks. AAAI 1 (2021), 196–203

    work page 2021

  20. [20]

    Junjie Huang, Huawei Shen, Liang Hou, and Xueqi Cheng. 2021. SDGNN: Learn- ing node representation for signed directed networks. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 35. 196–203

    work page 2021

  21. [21]

    Keke Huang, Jing Tang, Juncheng Liu, Renchi Yang, and Xiaokui Xiao. 2023. Node-wise diffusion for scalable graph learning. In Proceedings of the ACM web conference 2023. 1723–1733

    work page 2023

  22. [22]

    Wentao Huang, Yuchen Li, Yuan Fang, Ju Fan, and Hongxia Yang. 2020. BiANE: Bipartite Attributed Network Embedding. SIGIR 10, 20 (2020), 149–158

    work page 2020

  23. [23]

    Aditya Prakash, and Naren Ramakrishnan

    Mohammad Raihanul Islam, B. Aditya Prakash, and Naren Ramakrishnan. 2018. SIGNet: Scalable embeddings for signed networks. In Lect. Notes Comput. Sci. (including Subser. Lect. Notes Artif. Intell. Lect. Notes Bioinformatics) , Vol. 10938 LNAI. Springer-Verlag, Berlin, Heidelberg, 157–169. https://doi.org/10.1007/978- 3-319-93037-4_13

    doi:10.1007/978- 2018

  24. [24]

    Amin Javari, Tyler Derr, Pouya Esmailian, Jiliang Tang, and Kevin Chen-Chuan Chang. 2020. ROSE: Role-based Signed Network Embedding. In TheWebConf. 2782–2788

    work page 2020

  25. [25]

    Glen Jeh and Jennifer Widom. 2003. Scaling personalized web search. In Proceed- ings of the 12th international conference on World Wide Web . 271–279

    work page 2003

  26. [26]

    Jinhong Jung, Jaemin Yoo, and U Kang. 2022. Signed random walk diffusion for effective representation learning in signed graphs.Plos one 17, 3 (2022), e0265001

    work page 2022

  27. [27]

    Junghwan Kim, Haekyu Park, Ji-Eun Lee, and U Kang. 2018. Side: representation learning in signed directed networks. In Proceedings of the 2018 world wide web conference. 509–518

    work page 2018

  28. [28]

    Kipf and Max Welling

    Thomas N. Kipf and Max Welling. 2017. Semi-Supervised Classification with Graph Convolutional Networks. In International Conference on Learning Repre- sentations (ICLR)

    work page 2017

  29. [29]

    Johannes Klicpera, Aleksandar Bojchevski, and Stephan Günnemann. 2019. Pre- dict then propagate: Graph neural networks meet personalized PageRank. ICLR 2019 (oct 2019). arXiv:1810.05997 http://arxiv.org/abs/1810.05997

    Pith/arXiv arXiv 2019

  30. [30]

    Jérôme Kunegis and Andreas Lommatzsch. 2009. Learning spectral graph trans- formations for link prediction. In Proceedings of the 26th Annual International Conference on Machine Learning . 561–568

    work page 2009

  31. [31]

    Jure Leskovec, Daniel Huttenlocher, and Jon Kleinberg. 2010. Signed networks in social media. In Proceedings of the SIGCHI conference on human factors in computing systems. 1361–1370

    work page 2010

  32. [32]

    Yu Li, Meng Qu, Jian Tang, and Yi Chang. 2023. Signed laplacian graph neural networks. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 37. 4444–4452

    work page 2023

  33. [33]

    Yu Li, Yuan Tian, Jiawei Zhang, and Yi Chang. 2020. Learning signed network embedding via graph attention. AAAI 2020 - 34th AAAI Conf. Artif. Intell. (2020), 4772–4779. https://doi.org/10.1609/aaai.v34i04.5911

    work page doi:10.1609/aaai.v34i04.5911 2020

  34. [34]

    Yu Li, Yuan Tian, Jiawei Zhang, and Yi Chang. 2020. Learning signed network embedding via graph attention. In Proceedings of the AAAI conference on artificial intelligence, Vol. 34. 4772–4779

    work page 2020

  35. [35]

    David Liben-Nowell and Jon Kleinberg. 2003. The link prediction problem for so- cial networks. In Proceedings of the twelfth international conference on Information and knowledge management. 556–559

    work page 2003

  36. [36]

    László Lovász. 1993. Random walks on graphs. Combinatorics, Paul erdos is eighty 2, 1-46 (1993), 4

    work page 1993

  37. [37]

    Chunyu Lu, Pengfei Jiao, Hongtao Liu, Yaping Wang, Hongyan Xu, and Wenjun Wang. 2019. SSNE: Status Signed Network Embedding. In PAKDD. 81–93

    work page 2019

  38. [38]

    Linyuan Lü and Tao Zhou. 2011. Link prediction in complex networks: A survey. Physica A: statistical mechanics and its applications 390, 6 (2011), 1150–1170

    work page 2011

  39. [39]

    Paolo Massa and Paolo Avesani. 2007. Trust-aware recommender systems. In RecSys. 17–24

    work page 2007

  40. [40]

    Tomas Mikolov, Kai Chen, Greg Corrado, and Jeffrey Dean. 2013. Efficient esti- mation of word representations in vector space . 1–12 pages. arXiv:1301.3781

    Pith/arXiv arXiv 2013

  41. [41]

    Manfred Reimer. 2003. Multivariate Polynomial Approximation. Birkhäuser Basel. https://doi.org/10.1007/978-3-0348-8095-4

    work page doi:10.1007/978-3-0348-8095-4 2003

  42. [42]

    Fatemeh Salehi Rizi and Michael Granitzer. 2020. Signed heterogeneous network embedding in social media. In Proc. ACM Symp. Appl. Comput. 1877–1880. https: //doi.org/10.1145/3341105.3374048

    work page arXiv 2020

  43. [43]

    Yousef Saad. 2011. Numerical methods for large eigenvalue problems: revised edition. SIAM

    work page 2011

  44. [44]

    Jie Shen, Tao Tang, Li-Lian Wang, Jie Shen, Tao Tang, and Li-Lian Wang. 2011. Orthogonal polynomials and related approximation results. Spectral Methods: Algorithms, Analysis and Applications (2011), 47–140

    work page 2011

  45. [45]

    David I Shuman, Sunil K Narang, Pascal Frossard, Antonio Ortega, and Pierre Vandergheynst. 2013. The emerging field of signal processing on graphs: Ex- tending high-dimensional data analysis to networks and other irregular domains. IEEE signal processing magazine 30, 3 (2013), 83–98

    work page 2013

  46. [46]

    PK Suetin. 2001. Ultraspherical polynomials. Encyclopaedia of mathematics. Springer, Berlin (2001)

    work page 2001

  47. [47]

    Justin Sybrandt and Ilya Safro. 2020. FOBE and HOBE: First- and High-Order Bipartite Embeddings. arXiv:1905.10953 [cs.LG] https://arxiv.org/abs/1905.10953

    Pith/arXiv arXiv 2020

  48. [48]

    Hanzhi Wang, Mingguo He, Zhewei Wei, Sibo Wang, Ye Yuan, Xiaoyong Du, and Ji-Rong Wen. 2021. Approximate graph propagation. In Proceedings of the 27th ACM SIGKDD Conference on Knowledge Discovery & Data Mining . 1686–1696

    work page 2021

  49. [49]

    Hewen Wang, Renchi Yang, and Xiaokui Xiao. 2024. Effective Edge-wise Rep- resentation Learning in Edge-Attributed Bipartite Graphs. In KDD. 3081–3091. https://doi.org/10.1145/3637528.3671805

    arXiv 2024

  50. [50]

    Xiyuan Wang and Muhan Zhang. 2022. How Powerful are Spectral Graph Neural Networks. ICML 162 (2022), 23341–23362

    work page 2022

  51. [51]

    Weinberger

    Felix Wu, Tianyi Zhang, Amauri Holanda de Souza, Christopher Fifty, Tao Yu, and Kilian Q. Weinberger. 2019. Simplifying graph convolutional networks. In ICML, Vol. 2019-June. 11884–11894

    work page 2019

  52. [52]

    Xueyi Wu, Yuanyuan Xu, Wenjie Zhang, and Ying Zhang. 2023. Billion-Scale Bipartite Graph Embedding: A Global-Local Induced Approach. Proc. VLDB Endow. 17, 2 (2023), 175–183. https://doi.org/10.14778/3626292.3626300

    work page arXiv 2023

  53. [53]

    Hansheng Xue, Luwei Yang, Vaibhav Rajan, Wen Jiang, Yi Wei, and Yu Lin. 2021. Multiplex Bipartite Network Embedding Using Dual Hypergraph Convolutional Networks. In WWW. 1649–1660

    work page 2021

  54. [54]

    Renchi Yang and Jieming Shi. 2024. Efficient high-quality clustering for large bipartite graphs. Proceedings of the ACM on Management of Data 2, 1 (2024), 1–27. 10 GegenNet: Spectral Convolutional Neural Networks for Link Sign Prediction in Signed Bipartite Graphs CIKM ’25, November 10–14, 2025, Seoul, Republic of Korea

    work page 2024

  55. [55]

    Renchi Yang, Jieming Shi, Keke Huang, and Xiaokui Xiao. 2022. Scalable and Effective Bipartite Network Embedding. In Proceedings of the 2022 International Conference on Management of Data . 1977–1991

    work page 2022

  56. [56]

    Zeyu Zhang, Jiamou Liu, Kaiqi Zhao, Song Yang, Xianda Zheng, and Yifei Wang

    work page

  57. [57]

    Zeyu Zhang, Peiyao Zhao, Xin Li, Jiamou Liu, Xinrui Zhang, Junjie Huang, and Xiaofeng Zhu. 2024. Signed Graph Representation Learning: A Survey. arXiv preprint arXiv:2402.15980 (2024)

    Pith/arXiv arXiv 2024

  58. [58]

    Ziang Zhou, Jieming Shi, Renchi Yang, Yuanhang Zou, and Qing Li. 2023. SlotGAT: slot-based message passing for heterogeneous graphs. InInternational Conference on Machine Learning. PMLR, 42644–42657

    work page 2023

  59. [59]

    Meiqi Zhu, Xiao Wang, Chuan Shi, Houye Ji, and Peng Cui. 2021. Interpreting and unifying graph neural networks with an optimization framework. WWW (2021), 1215–1226. 11

    work page 2021

  60. [2023]

    In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval

    Contrastive Learning for Signed Bipartite Graphs. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval. 1629–1638

    work page

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.