Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

Information theory for hypergraph similarity

T0 review · 3 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A single normalized mutual information formula, equipped with bulk, aligned-layer, and cross-layer encodings, gives parameter-free hypergraph similarity measures that detect shared higher-order structure while suppressing spurious overlap f

desk verdict A solid, genuinely new methods paper on hypergraph similarity; the cross-order NMI is real progress, but the uniform-prior null model and missing baseline comparisons leave the strongest claims only partly supported. read the letter →

arxiv 2510.27411 v2 pith:3BRUCZHR submitted 2025-10-31 physics.soc-ph physics.data-an

Helcio Felippe , Alec Kirkley , Federico Battiston This is my paper
classification physics.soc-phphysics.data-an
keywords hypergraphsimilaritynormalizedmutualinformationhigher-orderinteractionsnetworkcomparisonminimumdescriptionlengthcross-ordermultiscalemeasurestheory
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

Most network comparison tools only see pairwise links, yet many real systems are organized by group interactions of three or more actors. This paper tries to establish a general, parameter-free way to measure similarity between hypergraphs: take any notion of structural overlap, encode it as an information-transmission scheme, and score the pair by how much knowing one hypergraph compresses the other. Three encodings are derived — bulk, layer-aligned, and cross-order — each capturing more granular kinds of correspondence, with the cross-order version able to match interactions across different sizes. If the construction is right, researchers get interpretable similarity scores bounded between 0 and 1 (1 for identical hypergraphs, near 0 for independent sparse ones), along with a multiscale extension that compares communities rather than individual nodes. The paper validates the measures on synthetic hypergraphs with planted overlap and applies them to collaboration data in physics, film, and software.

What carries the argument

The load-bearing object is an encoding-specific entropy Hc(G)=log(number of possible hypergraphs compatible with the encoding's constraints), paired with the conditional entropy Hc(G'|G) that counts configurations of G' after fixing its overlap with G. Mutual information is the bits saved, and the NMI is the larger of the two normalized savings directions (Eq. 4). The three encodings — bulk, align, cross — differ only in what is treated as known and how overlap is counted: align counts overlap within each edge-size layer separately, while cross counts overlap between a layer and the projections of higher-order layers, choosing for each target order the source layer that minimizes conditional

What would settle it

Build two independent random hypergraphs with the same planted node-degree sequences and per-layer edge counts, using a hypergraph configuration model, and compute NMIalign and NMIcross. If the scores do not fall to near zero for large N, the central 'correcting for spurious correlations' claim fails for degree-heterogeneous data; alternatively, if NMI scores between a real hypergraph and its degree-preserving randomized surrogates remain high, the null model is too weak.

Watch

Extended reading notes

Core claim

The central claim is that Eq. (4) — NMIc(G1,G2) = max(MIc(G1;G2)/Hc(G2), MIc(G2;G1)/Hc(G1)) — converts any encoding of hypergraph structure into a valid, normalized similarity measure, and that three specific encodings form a useful hierarchy. Under the bulk encoding all hyperedges are transmitted at once; under the align encoding each layer (edge size) is transmitted separately, which removes the artificial inflation that arises when layer densities differ; under the cross encoding a layer of one hypergraph can be transmitted from projected higher-order layers of the other, selecting the cheapest source layer for each target order, which captures nested and cross-order correspondence. The a

Load-bearing premise

All entropy counts assume that, given only the number of hyperedges in each layer, every possible hypergraph is equally likely; under that uniform null model any shared edge is treated as informative, so overlap produced by hubs or other degree heterogeneity is counted as similarity rather than spurious correlation.

Editorial extensions

If this is right

  • NMIcross can detect similarity between hypergraphs whose shared structure hides in different interaction orders, e.g., a triplet in one matching a pair in another via containment — something pairwise projection methods miss.
  • NMIalign removes the layer-density artifact that inflates NMIbulk, so it gives near-zero scores for independent hypergraphs even when sparser high-order layers make random overlap unlikely.
  • The measures are parameter-free, symmetric, bounded in [0,1], equal to 1 for identical hypergraphs, and the framework guarantees non-negativity for the proposed encodings.
  • The multiscale variant NMI(b)cross identifies shared community structure between hypergraphs generated from the same planted partition, with scores decaying smoothly as the community strength weakens or the partitions are shuffled.
  • A recursive overlap-counting algorithm avoids full projection of large hyperedges, so the cross measure scales to large hypergraphs despite the combinatorial explosion of sub-tuples.

Reading between the lines

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

  • Beyond the paper: because the framework accepts any encoding, a degree-corrected version that conditions on each node's hyperdegree sequence would separate true correspondence from hub-driven overlap — the paper corrects density heterogeneity but not degree heterogeneity, which it flags as future work.
  • Beyond the paper: at the extreme B=N in the coarse-graining, the multiscale measure compares multisets, so the same formulas apply to weighted or integer-valued hypergraph comparison without additional machinery.
  • Beyond the paper: the authors intentionally omit the cost of transmitting the overlap itself; charging that cost would make finite-size scores less inflated but would sacrifice the guaranteed non-negativity, a trade-off a follow-up could quantify.
  • Beyond the paper: the temporal anomaly application suggests NMIcross can serve as a change-point statistic for higher-order interaction streams, not just a static comparator between two snapshot hypergraphs.
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

3 major / 5 minor

Summary. The paper develops a general information-theoretic framework for hypergraph similarity. Given two simple hypergraphs on the same aligned node set, the authors define entropy and conditional entropy under a uniform prior over hypergraphs with fixed per-layer edge counts, and construct a normalized mutual information measure (Eq. 4). Three encodings are introduced: “bulk” (all hyperedges transmitted together), “align” (per-layer transmission with same-order overlaps), and “cross” (per-layer transmission allowing a layer of one hypergraph to condition a lower-order layer of the other, with a min over source layers). A multiscale extension replaces the node set by a node partition b and compares coarse-grained multisets (SM S3). The measures are tested on synthetic hypergraphs with injected noise and applied to three empirical multiplex higher-order networks (APS physics fields, IMDb movie genres, GitHub/Rust repositories), with results summarized as NMI matrices, dendrograms, and minimum spanning trees.

Significance. If the claims are correct, the framework is a substantial addition to the sparse toolkit for hypergraph comparison. Its strengths are the transparent combinatorial definitions, the absence of fitted parameters in the core bulk/align/cross measures, the explicit recursive algorithm for large-order projections (SM S2), and the public code. The cross encoding is a principled way to capture nested higher-order overlap, and the multiscale extension addresses a genuine need for community-level comparison. However, the validation is entirely internal: there is no comparison with existing hypergraph similarity methods, and the null model is a uniform prior over edge sets with fixed densities, which does not account for degree heterogeneity. The paper’s abstract claims the measures correct for “spurious correlations,” but the demonstrated correction is only for layer-density imbalances; extension to degree heterogeneity is explicitly deferred to future work. In addition, the multiscale conditional-entropy combinatorics appear to contain a technical error that affects the mesoscale experiments in Sec. III C. These issues make the central claims broader than what is currently established.

major comments (3)
  1. [Sec. II A, Eqs. (6), (9), (15); Sec. III; Discussion] All encodings use a uniform prior over hypergraphs with fixed per-layer edge counts. This null model treats every shared hyperedge as informative, so two independent degree-heterogeneous hypergraphs can receive inflated NMI scores because both are anchored by high-degree nodes. The synthetic experiments (Sec. III A and C) generate hyperedges uniformly at random and never exercise this regime. The Discussion explicitly defers degree-corrected variants to future work, so the abstract’s claim of “correcting for spurious correlations” is broader than what is demonstrated: the align and cross encodings correct for density heterogeneity across layers, but not for degree heterogeneity. Since the empirical hypergraphs in Fig. 5 have heavy-tailed degree distributions, the NMI values and the resulting dendrograms/MSTs may be partially driven by degree-driven edge overlap. Please either add a degre
  2. [SM S3, Eqs. (S11), (S13), (S15)] The multiscale measures are based on multisets of coarse-grained hyperedge types, and the entropy counts use the multiset coefficient. However, the conditional entropies count the number of ways to choose the overlap E^(b)_12 from the E_i coarse-grained hyperedges as the binomial coefficient C(E_i, E^(b)_12). For a multiset with repeated types, the number of possible overlap multisets is not a binomial coefficient; it is the number of integer vectors (n_x) with 0 ≤ n_x ≤ multiplicity(x) and sum n_x = E^(b)_12. This affects the multiscale NMI used in Sec. III C and Fig. 4. Please correct the combinatorics or clarify whether the encoding treats each coarse-grained edge as labeled, which would be inconsistent with the multiset entropy in Eq. (S9).
  3. [Sec. III B, Eq. (14)] NMI_cross chooses, for each target layer ℓ, the source layer k ≥ ℓ that minimizes the conditional entropy (Eq. 14). This minimization over many candidate layers can inflate the apparent similarity for independent hypergraphs, particularly when a high-order layer projects to a lower-order layer with very few target edges, making the conditional entropy term small. The paper does not report a null test for NMI_cross on pairs of independent random hypergraphs without planted cross-order structure. The experiments in Fig. 3 and SM S4 all contain some cross-order dependency by construction, so it is unclear whether NMI_cross approaches zero for uncorrelated hypergraphs. Please add a null experiment analogous to the ε = 1 case of Fig. 2 for NMI_cross with heterogeneous layer densities and no nestedness, and discuss the effect of the min in Eq. (14).
minor comments (5)
  1. [Sec. II A] The statement that 0 ≤ MI_c ≤ H_c holds for any encoding c is too general; SM S3 later acknowledges that for the multiscale bulk variant the conditional entropy can exceed the entropy, so a cutoff is needed. Please qualify the general claim in the main text.
  2. [Sec. III B, Fig. 3] The description that NMI_cross shows “a modest drop” in panel (d) is vague; please report the numerical values (and error bars) for both measures in all four columns.
  3. [SM S3, SM S5] There are several typos: “mutliscale” should be “multiscale”; “IMBd” and “IMDd” should be “IMDb”; “subbing” in the main text is nonstandard.
  4. [Sec. III D] The empirical applications are qualitative. Since the paper introduces a new similarity measure, a comparison with at least one existing hypergraph similarity method (e.g., Refs. [26] or [28]) on the same data would substantially strengthen the validation.
  5. [SM S3, Sec. III C] The multiscale measure requires a node partition b. If b is obtained from a community-detection algorithm, the method inherits that algorithm’s free parameters (and the description length of b is not included). Please clarify the status of b as input metadata versus an inferred quantity.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the hypergraph NMI measures are self-contained combinatorial definitions, validated on planted-overlap synthetic tests, and self-citations are not load-bearing.

full rationale

The paper's derivation chain is definitional in the normal, non-circular sense. Entropy and conditional entropy are defined as logarithms of explicit combinatorial counts (Eqs. 1–2, 6, 8, 9–10, 13–16), and the normalized mutual information in Eq. (4) is just a rearrangement of those counts. The properties highlighted by the paper—NMI = 1 for identical hypergraphs and NMI ≈ 0 for independent sparse random hypergraphs—follow from the counting formulas rather than from any fitted parameter or tuned constant. The synthetic experiments plant known overlaps and confirm that the scores respond monotonically to perturbation; this is standard validation of a new measure, not a prediction extracted from the data. The cross-order encoding's choice of the best source layer in Eq. (14) is an explicit modeling decision (MDL-style minimization over possible transmission layers), not a parameter fitted to make the experiments come out. The paper's self-citations ([47], [68]) provide the graph-level NMI that the hypergraph construction generalizes and state the convention of ignoring the codelength of the overlap, but the relevant graph formulas are re-derived in SM S1, so these citations are not load-bearing. The acknowledged limitation regarding degree heterogeneity (Discussion: a degree-corrected variant is left for future work) is a correctness/null-model concern about the uniform prior, not a circularity: no step in the derivation is equivalent by construction to its own input. Thus the circularity score is 0.

Assumptions & free parameters 1 free parameters · 6 assumptions · 0 invented entities

All measures rely on a uniform-prior null model over edge sets; there are no fitted parameters, but the multiscale similarity depends on a user-chosen partition. The key domain assumptions are aligned labeled nodes and one-way (large-to-small) cross-order projection.

free parameters (1)
  • node partition b for multiscale NMI = user-specified (B groups)
    SM S3: the multiscale measures NMI^(b) compare coarse-grained hypergraphs under a partition b, which may come from metadata or community detection; scores depend on this choice, though it is not fitted to data.
assumptions (6)
  • domain assumption Uniform prior over hypergraphs with fixed per-layer edge counts
    Invoked to write H_bulk (Eq. 6), H_align (Eq. 9), and H_cross (Eq. 15); all configurations with the same edge counts are assigned equal codelength. This is the null model against which similarity is measured.
  • domain assumption Simple, unweighted hypergraphs on the same N labeled nodes
    Sec. II A: the measures require aligned node labels and assume no multi- or self-edges, limiting direct application to unaligned or different-size systems.
  • domain assumption Cross-order overlap is restricted to k >= l (larger to smaller orders)
    Sec. II B: 'higher-order interactions cannot be uniquely determined from lower-order interactions alone' [54]; this defines what cross-order similarity can mean in the framework.
  • standard math Kraft inequality and prefix-code entropy
    SM S1: justifies interpreting log-counts as codelengths and entropy for valid probability distributions over graph configurations.
  • standard math Vandermonde identity for MI nonnegativity
    SM S1, Eq. (S5): used to show MI >= 0, i.e., conditioning never increases the combinatorial count.
  • standard math Multiset coefficient and multiset intersection for coarse-grained variants
    SM S3: needed to count configurations of coarse-grained hypergraphs with repeated group labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Information theory for hypergraph similarity." pith.science (2026). https://pith.science/paper/3BRUCZHR

@misc{pith2026251027411,
  author       = {Pith},
  title        = {Pith review of: Information theory for hypergraph similarity},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3BRUCZHR}},
  note         = {Machine review of arXiv:2510.27411}
}
read the original abstract

Comparing networks is essential for a number of downstream tasks, from clustering to anomaly detection. Despite higher-order interactions being critical for understanding the dynamics of complex systems, traditional approaches for network comparison are limited to pairwise interactions only. Here we construct a general information theoretic framework for hypergraph similarity, capturing meaningful correspondence among higher-order interactions while correcting for spurious correlations. Our method operationalizes any notion of structural overlap among hypergraphs as a principled normalized mutual information measure, allowing us to derive a hierarchy of increasingly granular formulations of similarity among hypergraphs within and across orders of interactions, and at multiple scales. We validate these measures through extensive experiments on synthetic hypergraphs and apply the framework to reveal meaningful patterns in a variety of empirical higher-order networks. Our work provides foundational tools for the principled comparison of higher-order networks, shedding light on the structural organization of networked systems with non-dyadic interactions.

Figures

Figures reproduced from arXiv: 2510.27411 by the authors.

Figure 1
Figure 1. FIG. 1. Hierarchy of information-theoretic measures for hypergraph similarity. Hypergraphs [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. FIG. 2. Information theory measures for intra-order hypergraph similarity. (a) Random hypergraphs with homogeneous layer [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗
Figure 3
Figure 3. FIG. 3. Information theory measures for cross-order hyper [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: FIG. 4. Mesoscale similarity for hypergraphs. (a) NMI [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]
Figure 5
Figure 5. Figure 5: FIG. 5. Hypergraph similarity for real-world systems. NMI matrices among all pairs of hypergraphs within real-world systems [PITH_FULL_IMAGE:figures/full_fig_p010_5.png]

Discussion (0). Sign in 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. Hypergraph backboning

    cs.SI 2026-05 unverdicted novelty 7.0 of 10

    Presents a non-parametric information-theoretic backboning method for hypergraphs that removes nested/redundant structures to yield a minimal representation of higher-order interactions, with extension to weighted cases.

Reference graph

Works this paper leans on

77 extracted references · 1 linked inside Pith · cited by 1 Pith paper

  1. [1]

    , B}at random to form the majority affiliation of a new hyperedge

    Choose a groupr∈ {1, . . . , B}at random to form the majority affiliation of a new hyperedge

  2. [2]

    Generate group affiliations for each of the remain- ingℓ−1 nodes by picking grouprwith probabil- itypand another community labels̸=runiformly at random from the remaining labels with proba- bility 1−p

  3. [3]

    detectability transition

    For each community label in the hyperedge, pick a node from that community uniformly at random, without replacement. This process results in random hypergraphs in which the expected fraction of nodes belonging to the major- ity community in each hyperedge isp. In this way, the individual hyperedges are independent and randomized acrossG 1, G2, so we would...

  4. [4]

    De Domenico, V

    M. De Domenico, V. Nicosia, A. Arenas, and V. La- tora, Structural reducibility of multilayer networks.Na- ture communications6(1), 1–9 (2015)

    2015

  5. [5]

    Ok, A graph similarity for deep learning.Advances in Neural Information Processing Systems33, 1–12 (2020)

    S. Ok, A graph similarity for deep learning.Advances in Neural Information Processing Systems33, 1–12 (2020)

    2020

  6. [6]

    Attar and S

    N. Attar and S. Aliakbary, Classification of complex net- works based on similarity of topological network features. Chaos: An interdisciplinary journal of nonlinear science 27(9) (2017)

    2017

  7. [7]

    Wills and F

    P. Wills and F. G. Meyer, Metrics for graph comparison: a practitioner’s guide.Plos One15(2), e0228728 (2020)

    2020

  8. [8]

    N. M. Kriege, F. D. Johansson, and C. Morris, A sur- vey on graph kernels.Applied Network Science5, 1–42 (2020)

    2020

Show all 77 references
  1. [9]

    Sharan and T

    R. Sharan and T. Ideker, Modeling cellular machinery through biological network comparison.Nature Biotech- nology24(4), 427–433 (2006)

    2006

  2. [10]

    Nikolova and J

    N. Nikolova and J. Jaworska, Approaches to measure chemical similarity–a review.QSAR & Combinatorial Science22(9-10), 1006–1026 (2003)

    2003

  3. [11]

    Mheich, F

    A. Mheich, F. Wendling, and M. Hassan, Brain network similarity: methods and applications.Network Neuro- science4(3), 507–527 (2020)

    2020

  4. [12]

    Faust and J

    K. Faust and J. Skvoretz, Comparing networks across space and time, size and species.Sociological Methodology 32(1), 267–299 (2002)

    2002

  5. [13]

    Lambiotte, M

    R. Lambiotte, M. Rosvall, and I. Scholtes, From networks to optimal higher-order models of complex systems.Na- ture physics15(4), 313–320 (2019)

    2019

  6. [14]

    Battiston, G

    F. Battiston, G. Cencetti, I. Iacopini, V. Latora, M. Lu- cas, A. Patania, J.-G. Young, and G. Petri, Networks beyond pairwise interactions: structure and dynamics. Physics Reports874, 1–92 (2020)

    2020

  7. [15]

    Battiston, E

    F. Battiston, E. Amico, A. Barrat, G. Bianconi, G. Fer- raz de Arruda, B. Franceschiello, I. Iacopini, S. K´ efi, V. Latora, Y. Moreno,et al., The physics of higher-order interactions in complex systems.Nature Physics17(10), 1093–1098 (2021)

    2021

  8. [16]

    Bianconi,Higher-order networks

    G. Bianconi,Higher-order networks. Cambridge Univer- sity Press (2021)

    2021

  9. [17]

    C. Bick, E. Gross, H. A. Harrington, and M. T. Schaub, What are higher-order networks?SIAM Review65(3), 686–731 (2023)

    2023

  10. [18]

    Berge,Hypergraphs: Combinatorics of Finite Sets, volume 45

    C. Berge,Hypergraphs: Combinatorics of Finite Sets, volume 45. Elsevier (1984)

    1984

  11. [19]

    Iacopini, G

    I. Iacopini, G. Petri, A. Barrat, and V. Latora, Simpli- cial models of social contagion.Nature Communications 10(1), 2485 (2019)

    2019

  12. [20]

    Neuh¨ auser, A

    L. Neuh¨ auser, A. Mellor, and R. Lambiotte, Multi- body interactions and nonlinear consensus dynamics on networked systems.Physical Review E101(3), 032310 (2020)

    2020

  13. [21]

    Ferraz de Arruda, A

    G. Ferraz de Arruda, A. Aleta, and Y. Moreno, Conta- gion dynamics on higher-order networks.Nature Reviews Physics6(8), 468–482 (2024)

    2024

  14. [22]

    Di Gaetano, G

    L. Di Gaetano, G. Carugno, F. Battiston, and F. Coghi, Dynamical fluctuations of random walks in higher- order networks.Physical Review Letters133(10), 107401 (2024)

    2024

  15. [23]

    A. P. Mill´ an, J. J. Torres, and G. Bianconi, Explo- sive higher-order kuramoto dynamics on simplicial com- plexes.Physical Review Letters124(21), 218301 (2020)

    2020

  16. [24]

    Zhang, M

    Y. Zhang, M. Lucas, and F. Battiston, Higher-order in- teractions shape collective dynamics differently in hy- pergraphs and simplicial complexes.Nature Communi- cations14(1), 1605 (2023)

    2023

  17. [25]

    Civilini, O

    A. Civilini, O. Sadekar, F. Battiston, J. G´ omez- Garde˜ nes, and V. Latora, Explosive cooperation in so- cial dilemmas on higher-order networks.Physical Review Letters132(16), 167401 (2024)

    2024

  18. [26]

    P. S. Chodrow, N. Veldt, and A. R. Benson, Generative hypergraph clustering: From blockmodels to modularity. Science Advances7(28), eabh1303 (2021)

    2021

  19. [27]

    Veldt, A

    N. Veldt, A. R. Benson, and J. Kleinberg, Combinato- rial characterizations and impossibilities for higher-order homophily.Science Advances9(1), eabq3200 (2023). 12

    2023

  20. [28]

    Ruggeri, M

    N. Ruggeri, M. Contisciani, F. Battiston, and C. De Bacco, Community detection in large hypergraphs. Science Advances9(28), eadg9159 (2023)

    2023

  21. [29]

    Surana, C

    A. Surana, C. Chen, and I. Rajapakse, Hypergraph sim- ilarity measures.IEEE Transactions on Network Science and Engineering(2022)

    2022

  22. [30]

    Martino and A

    A. Martino and A. Rizzi, (hyper) graph kernels over sim- plicial complexes.Entropy22(10), 1155 (2020)

    2020

  23. [31]

    Agostinelli, M

    C. Agostinelli, M. Mancastroppa, and A. Barrat, Higher- order dissimilarity measures for hypergraph comparison. arXiv preprint arXiv:2503.16959(2025)

    2025

  24. [32]

    S. Bai, F. Zhang, and P. H. Torr, Hypergraph convolu- tion and hypergraph attention.Pattern Recognition110, 107637 (2021)

    2021

  25. [33]

    R. Feng, T. Xu, X. Xie, Z.-K. Zhang, C. Liu, and X.-X. Zhan, A hyper-distance-based method for hypernetwork comparison.Chaos: An Interdisciplinary Journal of Non- linear Science34(8) (2024)

    2024

  26. [34]

    Saito, Hypergraph modeling via spectral embedding connection: Hypergraph cut, weighted kernel k-means, and heat kernel

    S. Saito, Hypergraph modeling via spectral embedding connection: Hypergraph cut, weighted kernel k-means, and heat kernel. InProceedings of the AAAI Confer- ence on Artificial Intelligence, volume 36, pp. 8141–8149 (2022)

    2022

  27. [35]

    L. Bai, E. R. Hancock, and P. Ren, A Jensen-Shannon kernel for hypergraphs. InStructural, Syntactic, and Statistical Pattern Recognition: Joint IAPR Interna- tional Workshop, SSPR&SPR 2012, Hiroshima, Japan, November 7-9, 2012. Proceedings, pp. 181–189, Springer (2012)

    2012

  28. [36]

    Lugo-Martinez, D

    J. Lugo-Martinez, D. Zeiberg, T. Gaudelet, N. Malod- Dognin, N. Przulj, and P. Radivojac, Classification in biological networks with hypergraphlet kernels.Bioinfor- matics37(7), 1000–1007 (2021)

    2021

  29. [37]

    D. J. MacKay,Information Theory, Inference and Learn- ing Algorithms. Cambridge University Press, Cambridge (2003)

    2003

  30. [38]

    Rissanen, Modeling by the shortest data description

    J. Rissanen, Modeling by the shortest data description. Automatica14, 465–471 (1978)

    1978

  31. [39]

    T. P. Peixoto, Bayesian stochastic blockmodeling. In P. Doreian, V. Batagelj, and A. Ferligoj (eds.),Advances in Network Clustering and Blockmodeling, pp. 289–332, Wiley, New York (2019)

    2019

  32. [40]

    Kirkley, Spatial regionalization based on optimal in- formation compression.Communications Physics5(1), 1–10 (2022)

    A. Kirkley, Spatial regionalization based on optimal in- formation compression.Communications Physics5(1), 1–10 (2022)

    2022

  33. [41]

    Morel-Balbi and A

    S. Morel-Balbi and A. Kirkley, Bayesian regionalization of urban mobility networks.Physical Review Research 6(3), 033307 (2024)

    2024

  34. [42]

    T. P. Peixoto, Reconstructing networks with unknown and heterogeneous errors.Phys. Rev. X8, 041011 (2018)

    2018

  35. [43]

    Kirkley, A

    A. Kirkley, A. Rojas, M. Rosvall, and J.-G. Young, Com- pressing network populations with modal networks reveal structural diversity.Communications Physics6(1), 148 (2023)

    2023

  36. [44]

    R. J. Gallagher, J.-G. Young, and B. F. Welles, A clar- ified typology of core-periphery structure in networks. Science Advances7(12), eabc9800 (2021)

    2021

  37. [45]

    Kirkley, Identifying hubs in directed networks.Phys- ical Review E109(3), 034310 (2024)

    A. Kirkley, Identifying hubs in directed networks.Phys- ical Review E109(3), 034310 (2024)

    2024

  38. [46]

    De Domenico and J

    M. De Domenico and J. Biamonte, Spectral entropies as information-theoretic tools for complex network compar- ison.Physical Review X6(4), 041062 (2016)

    2016

  39. [47]

    Escolano, E

    F. Escolano, E. R. Hancock, M. A. Lozano, and M. Cu- rado, The mutual information between graphs.Pattern Recognition Letters87, 12–19 (2017)

    2017

  40. [48]

    Corso, G

    G. Corso, G. M. Ferreira, and T. M. Lewinsohn, Mutual information as a general measure of structure in interac- tion networks.Entropy22(5), 528 (2020)

    2020

  41. [49]

    Coupette and J

    C. Coupette and J. Vreeken, Graph similarity descrip- tion: How are these graphs similar? InProceedings of the 27th ACM SIGKDD Conference on Knowledge Dis- covery & Data Mining, pp. 185–195 (2021)

    2021

  42. [50]

    Felippe, F

    H. Felippe, F. Battiston, and A. Kirkley, Network mutual information measures for graph similarity.Communica- tions Physics7(1), 335 (2024)

    2024

  43. [51]

    M. E. Newman, G. T. Cantwell, and J.-G. Young, Im- proved mutual information measure for clustering, clas- sification, and community detection.Physical Review E 101(4), 042304 (2020)

    2020

  44. [52]

    Jerdee, A

    M. Jerdee, A. Kirkley, and M. Newman, Mutual infor- mation and the encoding of contingency tables.Physical Review E110(6), 064306 (2024)

    2024

  45. [53]

    Q. F. Lotito, F. Musciotto, A. Montresor, and F. Battis- ton, Higher-order motif analysis in hypergraphs.Com- munications Physics5(1), 79 (2022)

    2022

  46. [54]

    LaRock and R

    T. LaRock and R. Lambiotte, Encapsulation structure and dynamics in hypergraphs.Journal of Physics: Com- plexity4(4), 045007 (2023)

    2023

  47. [55]

    N. W. Landry, J.-G. Young, and N. Eikmeier, The simpli- ciality of higher-order networks.EPJ Data Science13(1), 17 (2024)

    2024

  48. [56]

    Gallo, L

    L. Gallo, L. Lacasa, V. Latora, and F. Battiston, Higher- order correlations reveal complex memory in tempo- ral hypergraphs.Nature Communications15(1), 4754 (2024)

    2024

  49. [57]

    LaRock and R

    T. LaRock and R. Lambiotte, Exploring the non- uniqueness of node co-occurrence matrices of hyper- graphs.arXiv preprint arXiv:2506.01479(2025)

    2025 arXiv

  50. [58]

    Cencetti, F

    G. Cencetti, F. Battiston, B. Lepri, and M. Karsai, Tem- poral properties of higher-order interactions in social net- works.Scientific Reports11(1), 7028 (2021)

    2021

  51. [59]

    Decelle, F

    A. Decelle, F. Krzakala, C. Moore, and L. Zdeborov´ a, In- ference and phase transitions in the detection of modules in sparse networks.Physical Review Letters107, 065701 (2011). [57]https://journals.aps.org/datasets(2021)

    2011

  52. [60]

    Q. F. Lotito, A. Montresor, and F. Battiston, Multiplex measures for higher-order networks.Applied Network Sci- ence9(1), 55 (2024)

    2024

  53. [61]

    Schueller, J

    W. Schueller, J. Wachs, V. D. Servedio, S. Thurner, and V. Loreto, Evolving collaboration, dependencies, and use in the rust open source software ecosystem.Scientific data9(1), 703 (2022)

    2022

  54. [62]

    Schueller and J

    W. Schueller and J. Wachs, Modeling interconnected so- cial and technical risks in open source software ecosys- tems.Collective intelligence3(1), 26339137241231912 (2024)

    2024

  55. [63]

    Betti, L

    L. Betti, L. Gallo, J. Wachs, and F. Battiston, The dy- namics of leadership and success in software development teams.Nature Communications16(1), 1–11 (2025)

    2025

  56. [64]

    T. P. Peixoto and A. Kirkley, Implicit models, latent compression, intrinsic biases, and cheap lunches in com- munity detection.Physical Review E108(2), 024309 (2023)

    2023

  57. [65]

    Kirkley, Inference of dynamic hypergraph represen- tations in temporal interaction data.Physical Review E 109(5), 054306 (2024)

    A. Kirkley, Inference of dynamic hypergraph represen- tations in temporal interaction data.Physical Review E 109(5), 054306 (2024). 13

    2024

  58. [66]

    Santoro, F

    A. Santoro, F. Battiston, M. Lucas, G. Petri, and E. Am- ico, Higher-order connectomics of human brain function reveals local topological signatures of task decoding, in- dividual identification, and behavior.Nature Communi- cations15(1), 10244 (2024)

    2024

  59. [67]

    M. Neri, A. Brovelli, S. Castro, F. Fraisopi, M. Gatica, R. Herzog, P. A. Mediano, I. Mindlin, G. Petri, D. Bor, et al., A taxonomy of neuroscientific strategies based on interaction orders.European Journal of Neuroscience 61(3), e16676 (2025)

    2025

  60. [68]

    T. M. Cover and J. A. Thomas,Elements of Information Theory. John Wiley & Sons, Hoboken (2012)

    2012

  61. [69]

    F. S. Roberts and B. Tesman,Applied Combinatorics. CRC Press (2024)

    2024

  62. [70]

    Jerdee, A

    M. Jerdee, A. Kirkley, and M. Newman, Normalized mu- tual information is a biased measure for classification and community detection.arXiv preprint arXiv:2307.01282 (2023)

    2023

  63. [71]

    Ruggeri, F

    N. Ruggeri, F. Battiston, and C. De Bacco, Framework to generate hypergraphs with community structure.Physi- cal Review E109(3), 034309 (2024)

    2024

  64. [72]

    Iacopini, J

    I. Iacopini, J. R. Foote, N. H. Fefferman, E. P. Der- ryberry, and M. J. Silk, Not your private tˆ ete-` a-tˆ ete: leveraging the power of higher-order networks to study animal communication.Philosophical Transactions B 379(1905), 20230190 (2024)

    1905

  65. [73]

    Iacopini, M

    I. Iacopini, M. Karsai, and A. Barrat, The temporal dy- namics of group interactions in higher-order social net- works.Nature Communications15(1), 7391 (2024)

    2024

  66. [74]

    Klimt and Y

    B. Klimt and Y. Yang, The enron corpus: A new dataset for email classification research. InEuropean conference on machine learning, pp. 217–226, Springer (2004)

    2004

  67. [75]

    A. R. Benson, R. Abebe, M. T. Schaub, A. Jadbabaie, and J. Kleinberg, Simplicial closure and higher-order link prediction.Proceedings of the National Academy of Sci- ences115(48), E11221–E11230 (2018)

    2018

  68. [76]

    it’s al- ways about the people. enron is no different

    J. Diesner, T. L. Frantz, and K. M. Carley, Commu- nication networks from the enron email corpus “it’s al- ways about the people. enron is no different”.Computa- tional & Mathematical Organization Theory11(3), 201– 228 (2005)

    2005

  69. [77]

    zoom out

    J. Hardin, G. Sarkis, and P. Urc, Network analysis with the enron email corpus.Journal of Statistics Education 23(2) (2015). S1 Supplemental Material for: Information theory for hypergraph similarity Helcio Felippe,1 Alec Kirkley,1,2,3 and Federico Battiston4, 1Department of N...

    2015

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.