Pith. sign in

REVIEW 2 major objections 4 minor 1 cited by

Edge Correlations and Link Prediction in Growing Hypergraphs

T0 review · 2 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read One noisy-copy mechanism explains macroscopic statistics of temporal hypergraphs and yields an 11-parameter link predictor competitive with neural networks.

desk verdict Solid model and inference, but the intersection-size asymptotics have an algebraic slip that breaks the claimed intercepts as written. read the letter →

arxiv 2502.02386 v3 pith:ACCUDSHQ submitted 2025-02-04 cs.SI nlin.AOphysics.data-anphysics.soc-ph

Xie He , Philip S. Chodrow , Peter J. Mucha This is my paper · ORCID
classification cs.SInlin.AOphysics.data-anphysics.soc-ph MSC 05C6505C8005C82
keywords temporalhypergraphsgenerativemodeledgecopyinglinkpredictionstochasticexpectationmaximizationpower-lawdegreedistributionintersectionsizeshypergraphlikelihood
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that one mechanistic rule of hypergraph growth — each new edge is a noisy copy of a uniformly chosen earlier edge, supplemented with existing nodes drawn from $\gamma$ and brand-new nodes drawn from $\beta$ — is enough to reproduce several large-scale regularities of real temporal hypergraphs. If the claim is right, a complete hypergraph can be described by a low-dimensional likelihood rather than by subsampled sequences of edges, and future edges can be ranked by a probability the model assigns. The paper derives that edge sizes converge to the Perron eigenvector of a transition matrix $W$, that degree tails are power laws with an explicit exponent $\zeta$ in terms of $(\eta,\gamma,\beta)$, and that pairwise overlap densities obey a two-regime scaling: pairs sharing $k\ge 1$ nodes decay like $m^{-1}q_{ijk}$, while disjoint pairs approach constants $q_{ij0}$. Finally, it reports that a stochastic expectation-maximization fit of this model to empirical hypergraphs yields an 11-parameter link predictor competitive with much larger neural networks.

What carries the argument

The load-bearing object is the one-step update of Algorithm 1: uniform selection of an earlier edge, Bernoulli copying at rate $\eta$, addition of $g\sim\gamma$ extant nodes, and addition of $b\sim\beta$ novel nodes. Everything else in the paper is derived from this single update. The edge-size analysis reduces to the matrix $W$ (equation 3), whose closed-form binomial entries give the conditional distribution of the new edge size given the copied edge size, with the stationary edge-size distribution as its Perron eigenvector. The degree analysis reduces to a rate equation whose large-degree solution gives the power-law exponent $\zeta$ of equation (5). The intersection analysis reduces to a nonnegative matrix $C$ defined through a mean-field closure: the expected update of the pair count $P_{ijk}$ is closed by assuming the relevant conditional probabilities depend only on the sizes of the two edges and their intersection size. The Perron eigenvector $q=Cq$ then supplies the intercepts $q_{ijk}$ in the two-regime overlap scaling.

What would settle it

Simulate the HCM with fixed $\beta_0<1$ and $\eta<1$, and measure $r_{ijk}$ at two edge counts $m$ and $2m$. The claim predicts $r_{ij0}\to q_{ij0}$ and $m\,r_{ijk}\to q_{ijk}$ for $k\ge1$; if the ratio $m r_{ijk}$ does not stabilize or its limit differs from the Perron eigenvector of $C$ computed with an independent solver, equation (7) is falsified. A second check targets the reported large-$\eta$ failure: compute the spectrum of $C$ directly in that regime; if the spectral radius is not 1 or the leading eigenvector is not nonnegative, the stated relation between $q$ and $C$ fails there regardless of solver choice.

Watch

Extended reading notes

Core claim

On the paper's own terms, the discovery is that the Hyperedge Copy Model (HCM) is simultaneously a mechanistic growth process and an analytically tractable one. At each timestep the update picks an earlier edge $f$ uniformly, seeds the new edge with one node of $f$, keeps each other node of $f$ with probability $\eta$, then adds $g\sim\gamma$ existing nodes and $b\sim\beta$ novel nodes. The paper shows that the asymptotic edge-size distribution is the stationary distribution of the Markov chain with transition matrix $W$ (the Perron eigenvector), that $p_d\propto d^{-\zeta}$ with $\zeta = 1 + (1-\eta+\mu_\gamma+\mu_\beta)/(1-\eta(1-\mu_\gamma-\mu_\beta))$, and that the expected fraction of edge pairs with sizes $i,j$ and intersection size $k$ has the form $r_{ijk}=q_{ijk}+O(m^{-1})$ for $k=0$ and $r_{ijk}=m^{-1}q_{ijk}+O(m^{-2})$ for $k\ge1$, with $q$ the Perron eigenvector of a nonnegative matrix $C$ built from the model parameters. The same likelihood is used for inference, and the paper reports that fitting it with stochastic EM to 27 empirical hypergraphs gives parameters whose predicted degree and edge-size distributions broadly match the data, preserves non-vanishing edge overlap, and achieves link-prediction AUC and F1 scores competitive with neural baselines on several datasets.

Load-bearing premise

The load-bearing premise is an approximation: a pair of edges is assumed to be described completely by its two sizes and their overlap size, so deeper correlations can be ignored when writing the bookkeeping equations for edge pairs. The paper itself reports that the eigenproblem derived from this closure does not solve numerically for every parameter combination (large copying rates $\eta$), so the closure is asserted rather than fully proved.

Editorial extensions

If this is right

  • Given fitted parameters, the model outputs closed-form predictions for edge-size distribution, degree-tail exponent, and overlap densities, so mismatches with observed curves identify exactly which structural features the copying mechanism misses.
  • Because the model defines a likelihood over the complete hypergraph, data sets can be compared by likelihood and by the fitted copying rate and addition distributions without subsampling into edge sequences.
  • The two-regime overlap scaling means that measured overlap densities can serve as a fingerprint of the copying mechanism: the intercepts $q_{ijk}$ are set by the Perron eigenvector of $C$, not by ad-hoc mixing parameters.
  • A temporal hypergraph can be fitted through randomized pseudo-orderings, and the paper reports that this non-temporal fitting is often as predictive as fitting with true timestamps, so the model's predictive signal is largely carried by set-level copying rather than chronology.

Reading between the lines

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

  • Because $q$ is a deterministic function of $(\eta,\gamma,\beta)$, the overlap statistics could be used as a moment-based estimator: fit $q$ to empirical overlap rates and invert the eigenproblem, avoiding EM entirely; the paper does not develop this route.
  • The paper's suggested extension to recombination from multiple prior edges would break the single-parent structure that makes $q$ a Perron eigenvector, so the closed-form intersection theory should be viewed as the tractable boundary of this model family.
  • The near-parity of temporal and random-order training suggests a testable extension: permute the order of edges within the training window only; if predictive performance is unchanged, chronological order is not the carrier of signal.
  • If the reported numerical failure of the eigenproblem at large $\eta$ reflects genuine spectral behavior rather than solver error, the analytic claim for high-copying regimes needs a different proof strategy or additional assumptions.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 4 minor

Summary. The paper proposes the Hyperedge Copy Model (HCM), a temporal hypergraph generative model in which each new edge is formed by uniformly selecting an existing edge, copying a subset of its nodes with probability eta, adding g extant nodes according to gamma, and adding b novel nodes according to beta. The authors derive three asymptotic predictions: the edge-size distribution as the Perron eigenvector of a matrix W (eq. 3), a power-law degree tail with exponent zeta (eq. 5 / C7), and pairwise intersection-size densities r_ijk satisfying r_ijk = q_ijk + O(m^{-1}) for k=0 and r_ijk = m^{-1} q_ijk + O(m^{-2}) for k>=1, where q is the Perron eigenvector of a parameter-dependent matrix C (eq. 7). They fit the model with stochastic expectation maximization to 27 empirical hypergraphs, show qualitative agreement with degree and edge-size distributions, and benchmark link prediction against the neural methods NHP and LHP on three datasets, reporting competitive AUC/F1 with 11 parameters.

Significance. If the asymptotic claims hold, the paper makes a useful contribution: a single mechanistic copying rule organizes several macroscopic hypergraph statistics, the predictions are parameter-free functions of the fitted parameters and are tested against independently simulated HCM data, the SEM procedure scales to millions of nodes and edges, and the 11-parameter model is competitive with much larger neural link predictors. The authors also release code for reproduction and are candid about the heuristic nature of the degree-tail derivation and about datasets whose features are not captured. These strengths make the paper potentially valuable for network science, but the central intersection-size derivation currently contains a concrete algebraic error and several unproved assumptions that must be addressed before the claims in eq. (7) can be accepted.

major comments (2)
  1. [Supplementary C3, eqs. (C38)-(C43), (C53), (C85)] The hypergeometric term t^(4) is mis-specified and its large-m reduction has an inverted factor. Since extant nodes are drawn from N(t)\f, the eligible nodes of edge g are g\f, of size j-h, not j-sigma; the correct probability is HyperGeometric(k-sigma; x, j-h, n-l) = C(j-h, k-sigma) C(n-l-j+h, x-k+sigma) / C(n-l, x). For k=sigma+1 this has leading term x(j-h)/(n-l) = x(j-h)/(mu_beta m), using n = mu_beta m, whereas eq. (C43) writes x(j-sigma) mu_beta m^{-1}. At leading order only the h=0 contribution of psi survives, which removes the j-h versus j-sigma discrepancy, but the mu_beta^{-1} versus mu_beta factor persists: it enters w^(2) in eq. (C53), psi in eq. (C57), q_{ij1} in eq. (C85), and therefore the matrix C defining q in eq. (C87) and the intercepts in eq. (7). Figure 2 uses beta uniform on {1,2}, so mu_beta=1.5 and its predicted intercepts are affected by this error. The m^{-1} scaling is not invalidated by the correction, but the formula for q as written is not derived correctly.
  2. [Supplementary C3, eqs. (C60)-(C87) and C4] The determination of lambda_k and the claim that q is the unique nonnegative Perron eigenvector of C rest on assumptions that are asserted rather than proved: q_{ijk}>0 for k <= i^j, lambda_k != 2, and a contradiction argument whose conclusion is that a certain block has spectral radius strictly less than 1. If q has zero entries on some blocks, or if lambda_k=2 occurs, the argument does not close. The authors themselves report in Appendix C4 that the leading eigenpair of C could not be computed reliably for some parameters, especially large eta, and that the system was not solved for all parameter combinations. Since eq. (7) is stated unconditionally for beta_0<1, the analytic core of the intersection-size prediction is not fully established. In addition, the compartmental closure in eqs. (C14), (C23), and (C27) assumes that z and y depend only on edge sizes and intersection sizes and discards O(m^{-2}) terms without bounds, so the main text's phrase in Sec. II.C describing 'exact asymptotic structure' overstates the status of the derivation. The scaling law may survive a corrected argument, but the constants should be presented as following from an approximate mean-field closure under explicit conditions, or proved under the stated assumptions.
minor comments (4)
  1. [Appendix C3, text following eq. (C10)] The sentence says the matrix C is a function of 'alpha, beta, and gamma', but the model parameters are eta, gamma, and beta; alpha is not a model parameter.
  2. [Algorithm 1 and Sec. II.A] The extant-node addition step samples g distinct nodes uniformly from N(t)\f without specifying behavior when g exceeds |N(t)\f|, which can occur at early times or for large g; a convention or assumption should be stated.
  3. [Table II caption] There is a typo in the hardware description: 'a a 11th Gen Intel i7-11800H' should read 'an 11th Gen Intel i7-11800H'.
  4. [Sec. III.B and Table II] The comparison to neural baselines uses the NHP and LHP numbers as reported in the LHP paper, and the authors state they took the higher of two reported NHP scores; this selection rule should be stated in the main text, not only in the table caption, because it affects the fairness of the comparison.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HCM asymptotics are derived from the model's own update rule rather than fitted or self-cited inputs.

full rationale

The central derived quantities (mean edge size eq. (2), edge-size distribution via W and its Perron eigenvector eq. (3), degree power-law exponent eq. (C7), and intersection-size asymptotics eq. (7)) are explicit functions of the model parameters eta, gamma, and beta, and follow from the generative update in Algorithm 1. The intersection-size constants qijk are obtained from an approximate compartmental balance equation, leading to the consistency system q = Cq; this is an asymptotic self-consistency condition derived from the model, not an input fitted to the data. No load-bearing uniqueness theorem or ansatz is imported from the authors' prior work: the self-citations in the paper (e.g., refs. [16] and [17]) are background context and do not carry the derivation. The empirical link prediction benchmark is out-of-sample (20% training / 80% evaluation), with comparison to NHP and LHP using their published results on the same datasets, so there is no fitted-input-called-prediction circularity. The in-sample degree and edge-size comparisons in Figure 5 are model-fit diagnostics rather than claimed predictions. The paper itself flags limitations in Appendix C4, including numerical solver failures for large eta and unproved positivity/spectral-radius assumptions in the intersection-size argument; the supplementary hypergeometric step at eq. (C43) also appears to contain a scaling issue. These are correctness and rigor concerns, not circularity: they do not make the derivation equivalent to its own inputs. Overall the derivation chain is self-contained, and the modest in-sample checks and non-load-bearing self-citations do not constitute circular reasoning.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The paper uses three fitted objects: eta, gamma, beta, each learned per dataset; their length is set to the maximum observed edge size, itself a modeling choice. The asymptotics assume stationarity, beta_0 < 1, and Perron-Frobenius structure, which are stated but not fully justified. No new empirical entities are invented.

free parameters (3)
  • eta = varies per dataset, e.g., 0.887 for email-enron
    Copy probability, a scalar in [0,1] fitted via SEM; it controls the edge-size mean and the power-law exponent, so it is a genuine free parameter.
  • gamma distribution = varies per dataset, e.g., mu_gamma = 0.083 for email-enron
    Probability vector over the number of extant nodes to add, length up to max edge size; fitted via SEM as sufficient-statistic estimates s3,l.
  • beta distribution = varies per dataset, e.g., mu_beta = 0.013 for email-enron
    Probability vector over the number of novel nodes to add, length up to max edge size; fitted via SEM as s4,l.
assumptions (5)
  • domain assumption Uniform edge selection and uniform node selection within the sampled edge (Algorithm 1 steps 1-2).
    The entire likelihood and the p(f | e) posterior in the SEM derivation assume the generating edge is chosen uniformly and the seed vertex uniformly within it.
  • domain assumption Stationarity of degree and edge-size distributions in the asymptotic derivations (used to write p_{d}(t+1) = p_d(t) and mean-field closure).
    The self-consistent equations (1), (C2) and the intersection asymptotics assume the process has reached a stationary regime; no convergence theorem is proved for eta near 1.
  • domain assumption beta_0 < 1 and n large enough so the number of extant nodes n - |f| is large.
    The intersection-size scaling (C10) explicitly assumes beta_0 < 1, and the hypergeometric approximations (C40)-(C43) rely on extensive n in the large-m limit.
  • standard math Perron-Frobenius and spectral-radius arguments for the matrix C in the intersection asymptotics.
    The claim that q is the Perron eigenvector of C uses the Perron-Frobenius theorem of the matrix C, and the contradiction branch assumes the spectral radius is strictly less than one.
  • domain assumption The rate-equation derivation follows the Mitzenmacher heuristic for power-law tails.
    The extension from the dyadic preferential attachment rate equation to the hypergraph degree compartmental equation is heuristic and not a rigorous theorem.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Edge Correlations and Link Prediction in Growing Hypergraphs." pith.science (2026). https://pith.science/paper/ACCUDSHQ

@misc{pith2026250202386,
  author       = {Pith},
  title        = {Pith review of: Edge Correlations and Link Prediction in Growing Hypergraphs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ACCUDSHQ}},
  note         = {Machine review of arXiv:2502.02386}
}
read the original abstract

We propose a generative model of temporally-evolving hypergraphs in which hyperedges form via noisy copying of previous hyperedges. Our proposed model reproduces several stylized facts from many empirical hypergraphs, is learnable from data, and defines a likelihood over a complete hypergraph rather than ego-based or other sub-hypergraphs. Analyzing our model, we derive descriptions of node degree, edge size, and edge intersection size distributions in terms of the model parameters. We also show several features of empirical hypergraphs which are and are not successfully captured by our model. We provide a scalable stochastic expectation maximization algorithm with which we can fit our model to hypergraph data sets with millions of nodes and edges. Finally, we assess our model on a hypergraph link prediction task, finding that an instantiation of our model with just 11 parameters can achieve competitive predictive performance with large neural networks.

Figures

Figures reproduced from arXiv: 2502.02386 by the authors.

Figure 1
Figure 1. FIG. 1. Schematic illustration of the edge generation process for our Hyperedge Copy Model (HCM). (a): Current [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Illustration of the edge intersection asymptotics [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Structural properties of the [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (6 more)
Figure 4
Figure 4. Figure 4: FIG. 4. (Best viewed in color). Visual summary of parameters obtained by SEM fits of our HCM to empirical [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Degree distributions and edge-size distributions for one synthetic HCM with 10 [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: FIG. 6. Estimation of parameters in two synthetic hypergraphs with 10k hyperedges each. The convergence criterion, [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: FIG. 7. Mean edge sizes and mean node degrees compared between fit HCM instances and empirical data. Note the [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]
Figure 8
Figure 8. Figure 8: FIG. 8. Hypergraph properties including top-2 degree assortativity, clustering coefficient, edit simpliciality, and face [PITH_FULL_IMAGE:figures/full_fig_p015_8.png]
Figure 9
Figure 9. Figure 9: FIG. 9. Structural properties of the [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Broad Spectrum Structure Discovery in Large-Scale Higher-Order Networks

    cs.SI 2025-05 conditional novelty 6.0 of 10

    A two-level latent-class Poisson tensor model learns both assortative and disassortative mesoscale structure in large hypergraphs and improves heldout link prediction over an assortative-only baseline.

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages · cited by 1 Pith paper

  1. [1]

    The end of theory: The data del- uge makes the scientific method obsolete

    Chris Anderson. The end of theory: The data del- uge makes the scientific method obsolete. Wired magazine, 16(7):16–07, 2008

    work page 2008

  2. [2]

    Random preferential attachment hypergraph

    Chen Avin, Zvi Lotker, Yinon Nahum, and David Peleg. Random preferential attachment hypergraph. In Proceedings of the 2019 IEEE/ACM Interna- tional Conference on Advances in Social Networks Analysis and Mining , pages 398–405, Vancouver British Columbia Canada, August 2019. ACM

    work page 2019

  3. [3]

    Weighted simplicial complexes and their representation power of higher-order network data and topology

    Federica Baccini, Filippo Geraci, and Ginestra Bianconi. Weighted simplicial complexes and their representation power of higher-order network data and topology. Physical Review E , 106(3):034319, September 2022

    work page 2022

  4. [4]

    Hypergraphs with node attributes: Structure and inference, 2023

    Anna Badalyan, Nicol` o Ruggeri, and Caterina De Bacco. Hypergraphs with node attributes: Structure and inference, 2023

    work page 2023

  5. [5]

    Emer- gence of Scaling in Random Networks

    Albert-L´ aszl´ o Barab´ asi and R´ eka Albert. Emer- gence of Scaling in Random Networks. Science, 12 286(5439):509–512, October 1999

    work page 1999

  6. [6]

    The physics of higher-order interactions in complex systems

    Federico Battiston, Enrico Amico, Alain Barrat, Ginestra Bianconi, Guilherme Ferraz de Arruda, Benedetta Franceschiello, Iacopo Iacopini, Sonia K´ efi, Vito Latora, Yamir Moreno, et al. The physics of higher-order interactions in complex systems. Na- ture Physics, 17(10):1093–1098, 2021

    work page 2021

  7. [7]

    Networks be- yond pairwise interactions: Structure and dynamics

    Federico Battiston, Giulia Cencetti, Iacopo Iacopini, Vito Latora, Maxime Lucas, Alice Patania, Jean- Gabriel Young, and Giovanni Petri. Networks be- yond pairwise interactions: Structure and dynamics. Physics Reports, 874:1–92, August 2020

    work page 2020

  8. [8]

    Benson, Rediet Abebe, Michael T

    Austin R. Benson, Rediet Abebe, Michael T. Schaub, Ali Jadbabaie, and Jon Kleinberg. Sim- plicial closure and higher-order link prediction. Proceedings of the National Academy of Sciences , 115(48):E11221–E11230, November 2018

    work page 2018

Show all 63 references
  1. [9]

    Benson, Ravi Kumar, and Andrew Tomkins

    Austin R. Benson, Ravi Kumar, and Andrew Tomkins. Sequences of Sets. In Proceedings of the 24th ACM SIGKDD International Conference on Knowledge Discovery & Data Mining , pages 1148– 1157, London United Kingdom, July 2018. ACM

    2018

  2. [10]

    Har- rington, and Michael T

    Christian Bick, Elizabeth Gross, Heather A. Har- rington, and Michael T. Schaub. What Are Higher- Order Networks? SIAM Review, 65(3):686–731, Au- gust 2023

    2023

  3. [11]

    On-Line Expectation–Maximization Algorithm for latent Data Models

    Olivier Capp´ e and Eric Moulines. On-Line Expectation–Maximization Algorithm for latent Data Models. Journal of the Royal Statistical Soci- ety Series B: Statistical Methodology, 71(3):593–613, June 2009

    2009

  4. [12]

    Temporal properties of higher- order interactions in social networks

    Giulia Cencetti, Federico Battiston, Bruno Lepri, and M´ arton Karsai. Temporal properties of higher- order interactions in social networks. Scientific Re- ports, 11(1):7028, March 2021

    2021

  5. [13]

    A survey on hyperlink prediction

    Can Chen and Yang-Yu Liu. A survey on hyperlink prediction. IEEE Transactions on Neural Networks and Learning Systems , 2023

    2023

  6. [14]

    Explainable link prediction in knowledge hyper- graphs

    Zirui Chen, Xin Wang, Chenxu Wang, and Jianxin Li. Explainable link prediction in knowledge hyper- graphs. In Proceedings of the 31st ACM interna- tional conference on information & knowledge man- agement, pages 262–271, 2022

    2022

  7. [15]

    Random walks on hypergraphs with edge-dependent ver- tex weights

    Uthsav Chitra and Benjamin Raphael. Random walks on hypergraphs with edge-dependent ver- tex weights. In Kamalika Chaudhuri and Ruslan Salakhutdinov, editors, Proceedings of the 36th In- ternational Conference on Machine Learning , vol- ume 97 of Proceedings of Machine Learnin...

    2019

  8. [16]

    Philip S. Chodrow. Configuration models of ran- dom hypergraphs. Journal of Complex Networks , 8(3):cnaa018, 2020

    2020

  9. [17]

    Chodrow and Andrew Mellor

    Philip S. Chodrow and Andrew Mellor. Annotated hypergraphs: Models and applications. Applied Net- work Science, 5(1):9, December 2020

    2020

  10. [18]

    Dempster, Nan M

    Arthur P. Dempster, Nan M. Laird, and Donald B. Rubin. Maximum likelihood from incomplete data via the EM algorithm. Journal of the Royal Statis- tical Society: Series B (Methodological) , 39(1):1–22, 1977

    1977

  11. [19]

    Goldberg

    Suzanne Renick Gallagher and Debra S. Goldberg. Clustering Coefficients in Protein Interaction Hy- pernetworks. In Proceedings of the International Conference on Bioinformatics, Computational Bi- ology and Biomedical Informatics , pages 552–560, Wshington DC USA, September 2013. ACM

    2013

  12. [20]

    Directed hypergraphs and appli- cations

    Giorgio Gallo, Giustino Longo, Stefano Pallottino, and Sang Nguyen. Directed hypergraphs and appli- cations. Discrete Applied Mathematics, 42(2-3):177– 201, 1993

    1993

  13. [21]

    Preferential attachment hypergraph with vertex deactivation, April 2022

    Fr´ ed´ eric Giroire, Nicolas Nisse, Kostiantyn Ohulchanskyi, Ma lgorzata Sulkowska, and Thibaud Trolliet. Preferential attachment hypergraph with vertex deactivation, April 2022

    2022

  14. [22]

    Preferential attachment hypergraph with high modularity

    Fr´ ed´ eric Giroire, Nicolas Nisse, Thibaud Trolliet, and Ma lgorzata Sulkowska. Preferential attachment hypergraph with high modularity. Network Science, 10(4):400–429, December 2022

    2022

  15. [23]

    Kleinberg, Ravi Kumar, Prabhakar Ragha- van, Sridhar Rajagopalan, and Andrew S

    Jon M. Kleinberg, Ravi Kumar, Prabhakar Ragha- van, Sridhar Rajagopalan, and Andrew S. Tomkins. The Web as a Graph: Measurements, Mod- els, and Methods. In G. Goos, J. Hartmanis, J. Van Leeuwen, Takano Asano, Hideki Imai, D. T. Lee, Shin-ichi Nakano, and Takeshi Tokuyama, ed- ...

    1999

  16. [24]

    Landry, Maxime Lucas, Iacopo Ia- copini, Giovanni Petri, Alice Schwarze, Alice Pata- nia, and Leo Torres

    Nicholas W. Landry, Maxime Lucas, Iacopo Ia- copini, Giovanni Petri, Alice Schwarze, Alice Pata- nia, and Leo Torres. XGI: A Python package for higher-order interaction networks. Journal of Open Source Software, 8(85):5162, May 2023

    2023

  17. [25]

    Landry, Jean-Gabriel Young, and Nicole Eikmeier

    Nicholas W. Landry, Jean-Gabriel Young, and Nicole Eikmeier. The simpliciality of higher-order networks. EPJ Data Science, 13(1):17, March 2024

    2024

  18. [26]

    A Survey on Hypergraph Mining: Pat- terns, Tools, and Generators, January 2024

    Geon Lee, Fanchen Bu, Tina Eliassi-Rad, and Ki- jung Shin. A Survey on Hypergraph Mining: Pat- terns, Tools, and Generators, January 2024

    2024

  19. [27]

    How Do Hyperedges Overlap in Real-World Hy- pergraphs? - Patterns, Measures, and Generators

    Geon Lee, Minyoung Choe, and Kijung Shin. How Do Hyperedges Overlap in Real-World Hy- pergraphs? - Patterns, Measures, and Generators. In Proceedings of the Web Conference 2021 , pages 3396–3407, Ljubljana Slovenia, April 2021. ACM

    2021

  20. [28]

    Hypergraph motifs: Concepts, algorithms, and discoveries

    Geon Lee, Jihoon Ko, and Kijung Shin. Hypergraph motifs: Concepts, algorithms, and discoveries. Pro- ceedings of the VLDB Endowment , 13(12):2256– 2269, August 2020

    2020

  21. [29]

    THyMe+: Temporal Hypergraph Motifs and Fast Algorithms for Exact Counting

    Geon Lee and Kijung Shin. THyMe+: Temporal Hypergraph Motifs and Fast Algorithms for Exact Counting. In 2021 IEEE International Conference on Data Mining (ICDM) , pages 310–319, Auckland, New Zealand, December 2021. IEEE

    2021

  22. [30]

    The link- prediction problem for social networks

    David Liben-Nowell and Jon Kleinberg. The link- prediction problem for social networks. Journal of the American Society for Information Science and Technology, 58(7):1019–1031, May 2007

    2007

  23. [31]

    Hypergraph reconstruction from uncer- tain pairwise observations

    Simon Lizotte, Jean-Gabriel Young, and Antoine Allard. Hypergraph reconstruction from uncer- tain pairwise observations. Scientific Reports , 13(1):21364, December 2023. 13

    2023

  24. [32]

    Higher- order motif analysis in hypergraphs

    Quintino Francesco Lotito, Federico Musciotto, Al- berto Montresor, and Federico Battiston. Higher- order motif analysis in hypergraphs. Communica- tions Physics , 5(1):79, April 2022

    2022

  25. [33]

    Link predic- tion using low-dimensional node embeddings: The measurement problem

    Nicolas Menand and C Seshadhri. Link predic- tion using low-dimensional node embeddings: The measurement problem. Proceedings of the National Academy of Sciences, 121(8):e2312527121, 2024

    2024

  26. [34]

    Link pre- diction in social networks using hyper-motif rep- resentation on hypergraph

    ChunYan Meng and Hooman Motevalli. Link pre- diction in social networks using hyper-motif rep- resentation on hypergraph. Multimedia Systems , 30(3):123, 2024

    2024

  27. [35]

    R. Milo, S. Shen-Orr, S. Itzkovitz, N. Kashtan, D. Chklovskii, and U. Alon. Network Motifs: Sim- ple Building Blocks of Complex Networks. Science, 298(5594):824–827, October 2002

    2002

  28. [36]

    A Brief History of Genera- tive Models for Power Law and Lognormal Distribu- tions

    Michael Mitzenmacher. A Brief History of Genera- tive Models for Power Law and Lognormal Distribu- tions. Internet Mathematics, 1(2):226–251, January 2004

    2004

  29. [37]

    Topological Analysis of Temporal Hypergraphs

    Audun Myers, Cliff Joslyn, Bill Kay, Emilie Purvine, Gregory Roek, and Madelyn Shapiro. Topological Analysis of Temporal Hypergraphs. In Megan Dewar, Pawe l Pra lat, Przemys law Szufel, Fran¸ cois Th´ eberge, and Ma lgorzata Wrzosek, edi- tors, Algorithms and Models for the We...

    2023

  30. [38]

    Leonie Neuh¨ auser, Renaud Lambiotte, and Michael T. Schaub. Consensus dynamics on tempo- ral hypergraphs. Physical Review E, 104(6):064305, December 2021

    2021

  31. [39]

    Mark E. J. Newman. Networks: An Introduction . Oxford University Press, 2018

    2018

  32. [40]

    Dahae Roh and K. I. Goh. Growing hypergraphs with preferential linking. Journal of the Korean Physical Society, 83(9):713–722, November 2023

    2023

  33. [41]

    Community de- tection in large hypergraphs

    Nicol` o Ruggeri, Martina Contisciani, Federico Bat- tiston, and Caterina De Bacco. Community de- tection in large hypergraphs. Science Advances , 9(28):eadg9159, July 2023

    2023

  34. [42]

    Modelling non-linear consensus dynamics on hypergraphs

    Rohit Sahasrabuddhe, Leonie Neuh¨ auser, and Re- naud Lambiotte. Modelling non-linear consensus dynamics on hypergraphs. Journal of Physics: Complexity, 2(2):025006, June 2021

    2021

  35. [43]

    Sol´ e, Romualdo Pastor-Satorras, Eric Smith, and Thomas B

    Ricard V. Sol´ e, Romualdo Pastor-Satorras, Eric Smith, and Thomas B. Kepler. A Model of Large- Scale Proteome Evolution. Advances in Complex Systems, 05(01):43–54, March 2002

    2002

  36. [44]

    Qi Suo, Jin-Li Guo, Shiwei Sun, and Han Liu. Ex- ploring the evolutionary mechanism of complex sup- ply chain systems using evolving hypergraphs.Phys- ica A: Statistical Mechanics and its Applications , 489:141–148, January 2018

    2018

  37. [45]

    Rxn Hyper- graph: A Hypergraph Attention Model for Chemical Reaction Representation, 2022

    Mohammadamin Tavakoli, Alexander Shmakov, Francesco Ceccarelli, and Pierre Baldi. Rxn Hyper- graph: A Hypergraph Attention Model for Chemical Reaction Representation, 2022

    2022

  38. [46]

    Blevins, Danielle Bassett, and Tina Eliassi-Rad

    Leo Torres, Ann S. Blevins, Danielle Bassett, and Tina Eliassi-Rad. The Why, How, and When of Representations for Complex Systems. SIAM Re- view, 63(3):435–485, January 2021

    2021

  39. [47]

    Modeling of Pro- tein Interaction Networks

    Alexei V´ azquez, Alessandro Flammini, Amos Mar- itan, and Alessandro Vespignani. Modeling of Pro- tein Interaction Networks. Complexus, 1(1):38–44, 2003

    2003

  40. [48]

    NHP: Neural Hypergraph Link Prediction

    Naganand Yadati, Vikram Nitin, Madhav Nimishakavi, Prateek Yadav, Anand Louis, and Partha Talukdar. NHP: Neural Hypergraph Link Prediction. In Proceedings of the 29th ACM International Conference on Information & Knowl- edge Management, pages 1705–1714, Virtual Event Ireland, ...

    2020

  41. [49]

    LHP: Logical hyper- graph link prediction

    Yang Yang, Xue Li, Yi Guan, Haotian Wang, Chao- ran Kong, and Jingchi Jiang. LHP: Logical hyper- graph link prediction. Expert Systems with Applica- tions, 222:119842, July 2023

    2023

  42. [50]

    Jean-Gabriel Young, Giovanni Petri, and Tiago P. Peixoto. Hypergraph reconstruction from network data. Communications Physics , 4(1):135, June 2021

    2021

  43. [51]

    A hypergraph model of social tagging networks

    Zi-Ke Zhang and Chuang Liu. A hypergraph model of social tagging networks. Journal of Statistical Me- chanics: Theory and Experiment , 2010(10):P10005, October 2010. 14 Appendix A: Parameter estimation in synthetic HCM hypergraphs. For consistency, we applied the same converge...

    2010

  44. [52]

    Our argument is based on rate equations, generalizing an argument by Mitzen- macher [36] in the context of preferential attachment dyadic graphs

    Power-law degree distribution We now give a heuristic argument that the de- gree distribution of our model has an asymptotic power-law tail, with an exponent that depends on the model parameters. Our argument is based on rate equations, generalizing an argument by Mitzen- mach...

    1927

  45. [53]

    Edge-size distribution We describe the entries of the matrixW described in the main text. Each entry wij of W represents the conditional probability wij = P (|e| = i | |f | = j) , (C8) where f is the edge sampled in the edge-selection step of Algorithm 1 and e is the final edg...

  46. [54]

    successful

    Asymptotic properties of intersection sizes We now develop a probabilistic argument support- ing the following claim: Claim. Let rijk be the expected proportion of pairs of the m edges, relative to the total number of pairsm 2 , which have sizes i and j with intersection size ...

  47. [55]

    So far, we have shown that, for anyk such that the first two terms of eq

    Therefore, we must have h∗ = 0, from which it follows that λk = 1. So far, we have shown that, for anyk such that the first two terms of eq. (C73) do not vanish, we must have λk = 1. We will now show that if the two terms of eq. (C73) do vanish for somek = k∗, then qijh = 0 fo...

  48. [56]

    Computational Challenges In the experiment shown in the main text, we tracked edge sizes and intersections up to size 12, resulting in a matrix C of size 123 ×123. Experimen- tally, we found that the LAPACK solver (accessed through the numpy.linalg.eig function in Python) was ...

  49. [57]

    Growing Hypergraph Erd˝ os-R´ enyi model We list the steps that we take for growing hy- pergraphs in an Erd˝ os-R´ enyi manner in a frame- work similar to what we did for HCM in the main text. We emphasize that our approach here is only one way of generating hypergraphs that g...

  50. [58]

    Extant node sampling: Following the above notation, we select α + g nodes drawn uni- formly at random without replacement from N (t) to initiate e(t+1)

  51. [59]

    After forming e(t+1), we have an Erd˝ os-R´ enyi update H(t+1) = (N (t) ∪ e(t+1) , E (t) ∪ e(t+1))

    Novel node addition: We next sample ˆb from a Poisson distribution with mean b (so that there is some randomness in this number compared to the HCM) and add ˆb novel nodes to e(t+1). After forming e(t+1), we have an Erd˝ os-R´ enyi update H(t+1) = (N (t) ∪ e(t+1) , E (t) ∪ e(t+1))

  52. [60]

    Hypergraph Preferential Attachment Model We do not employ the model of [2] due to lack of available code for simulation or inference. Instead, the generalization of preferential attachment we use here takes the different numbers from the HCM step, using the two different numbe...

  53. [61]

    Degree based sampling: We select α nodes drawn with probability proportional to their degree without replacement from N (t) and name this set α′

  54. [62]

    Extant node sampling: We select g nodes drawn uniformly at random without replace- ment from N (t) \ α′ to initiate e(t+1)

  55. [63]

    After forming e(t+1), we have a Preferential Attach- ment update H(t+1) = (N (t)∪ e(t+1) , E (t)∪e(t+1))

    Novel node addition: We next sample ˆb from a Poisson distribution with mean b (so that there is some randomness in this number compared to the HCM) and add ˆb novel nodes to e(t+1). After forming e(t+1), we have a Preferential Attach- ment update H(t+1) = (N (t)∪ e(t+1) , E (...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.