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

Aaron Schein; Caterina De Bacco; John Hood

REVIEW 3 major objections 4 minor 67 references

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

T0 review · 3 major / 4 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read A low-rank Poisson model that places classes inside a latent assortative hypergraph provably discovers the assortative-to-disassortative spectrum in large hypergraphs.

desk verdict The omniassortative E-step omits the community rates, so the paper's central empirical claims rest on an unverified algorithm; the model and identifiability theory are still worth taking seriously. read the letter →

arxiv 2505.21748 v1 pith:33JHEWCU submitted 2025-05-27 cs.SI stat.COstat.MEstat.ML

John Hood , Caterina De Bacco , Aaron Schein This is my paper · ORCID

classification cs.SIstat.COstat.MEstat.ML

keywords hypergraphshigher-ordernetworksstochasticblockmodelstensordecompositiondisassortativestructurecommunitydetectionPoissonfactorizationlinkprediction

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

The paper introduces Omni-Hype-SMT, a probabilistic generative model for hypergraphs that tries to represent the full spectrum of mesoscale structure, from strict communities to disassortative classes, without an exponential growth in parameters. It does this by factoring observed hyperedge rates into soft node-to-class memberships ($\Theta$) and a class-affinity tensor, and then factoring the class-affinity tensor itself as a sum over latent communities, so that classes interact only inside communities. The authors prove that under a separability constraint (each class has a pure community of its own), the model's parameters are uniquely identified up to permutation, and that a leading strictly assortative hypergraph model is a special case. Empirically, the model improves held-out link prediction on five of six real-world hypergraph datasets, produces sharper class memberships, and generates realistic synthetic hypergraphs quickly. A curious reader should care because this offers one tractable and interpretable model for discovering group structure in higher-order network data without knowings in advance whether that structure is assortative or disassortative.

What carries the argument

The carrying object is a latent hypergraph of classes: each node has a soft membership vector $\theta_i$ over $C$ classes, and each class has a soft membership vector $w_k$ over $K$ communities, with the first $C$ columns of $W$ set to the identity. The class-affinity tensor $\Lambda^{(d)}$ is modeled by a symmetric CP decomposition $\Lambda^{(d)}_{c_1\dots c_d} = \sum_k \gamma^{(d)}_k \prod_q w_{c_q k}$, which encodes the assumption that classes interact exclusively within communities. That assumption is what suppresses the combinatorial explosion of inter-class interaction parameters while still allowing nodes of different classes to interact disassortatively. Identifiability is carried by the pure-community constraint $W=[I_C\mid\cdots]$ together with the CP-uniqueness bound $K \le \tfrac12(d(C-1)+1)$, which lets the authors separate the class structure from the community structure in the estimated parameters.

What would settle it

Simulate hypergraph data from the same model family with maximum order $d \geq 3$ but with $W$ chosen so that no column is a pure class, thereby violating Eq. (5), then fit Omni-Hype-SMT from many random initializations and check whether the maximum-likelihood estimates collapse to a single orbit up to permutation; distinct fitted parameter sets giving the same predictive distribution would show that the advertised uniqueness claim fails once its key premise is removed.

Watch

Extended reading notes

Core claim

The central claim is that higher-order interactions among nodes can be explained by two nested latent clusterings: nodes into classes, and classes into communities. Writing the observed $d$-way interaction counts as a tensor $A^{(d)}$ with Poisson rates $\mu^{(d)}$, the model factorizes each rate as $\mu^{(d)}_{i_1\dots i_d} = \sum_{c_1,\dots,c_d} \Lambda^{(d)}_{c_1\dots c_d} \prod_{r=1}^d \theta_{i_r c_r}$, where $\Theta$ is the node-class membership matrix. The class-affinity tensor is then itself factorized as $\Lambda^{(d)}_{c_1\dots c_d} = \sum_{k=1}^K \gamma^{(d)}_k \prod_{q=1}^d w_{c_q k}$, with $W$ a class-community membership matrix and $\gamma^{(d)}_k$ per-order community rates. This makes the observed hypergraph omniassortative: strict assortativity, core-periphery, and disassortative patterns all appear as parameter regimes of one family. The paper's main theoretical burden is identifiability, proved by combining CP-uniqueness criteria with a separability result from nonnegative matrix factorization under the constraint $W = [I_C \mid \cdots]$, yielding unique recovery of $\Theta$, $W$, and $\Gamma$ up to permutation of classes and communities.

Load-bearing premise

The claim that fitted classes and communities have a unique meaning rests on the imposed assumption that each latent class has a pure community containing only that class; real data without such pure communities fall outside the identifiability guarantee.

Editorial extensions

If this is right

Researchers can fit a single model that spans assortative, core-periphery, and disassortative structure instead of committing to one restricted structure in advance.
Because parameters are identifiable up to permutation, inferred classes and communities can be interpreted as meaningful latent structure rather than arbitrary axes of a non-unique factorization.
The strictly assortative model in Ref. [6] is a special case, so Omni-Hype-SMT provides a common framework that contains existing hypergraph community-detection approaches.
The generative procedure lets practitioners sample synthetic hypergraphs with tunable mesoscale structure, and the paper demonstrates that the simulated data closely match real aggregate statistics.
Held-out likelihood gains tend to grow with hyperedge order, suggesting that modeling disassortative structure is more important for larger multi-way interactions.

Reading between the lines

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

The authors leave implicit that the pure-community constraint is a substantive modeling assumption: if a real dataset has no latent class that ever appears alone, the identifiability theorem does not cover the fitted solution, and external validation of the recovered clusters would be needed.
The same classes-as-nodes construction could plausibly be carried into directed or temporal hypergraphs, though the symmetry and identifiability arguments would have to be reworked, and the paper only sketches these as future directions.
A testable extension is to use the per-order rates $\gamma^{(d)}_k$ as a diagnostic for interaction-order-dependent behavior: stability of the rate ordering across random data splits would give a model-based check for whether a community's role changes with hyperedge size.

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.

Desk Editor's Note

read the letter

What you should know: this paper has a genuinely new two-level class-community factorization for hypergraphs, but the omniassortative inference algorithm in Algorithm 2 omits the community rates γ from the E-step. If the code matches the text, the headline empirical results are not MLEs of the stated model.

The model itself is a real extension. Treating node classes as nodes in a latent hypergraph and imposing a CP decomposition on the class-affinity tensor is a clever way to get disassortative structure without exponential parameters. The identifiability theorem under the separability constraint W=[I_C|...] is a legitimate contribution, and the fast synthetic hypergraph generator is a plus. The code is public and the heldout link-prediction setup is genuine out-of-sample evaluation.

The soft spot is load-bearing and concrete. Eq. (S17) and Algorithm 2 compute the expected community subcount with the bracket only, no γ^(d)_k factor, whereas the model's multinomial thinning (Eq. S11) says the probability of community k is proportional to γ^(d)_k times that bracket divided by γ^(d)_k' times the bracket for k'. Those rates do not cancel. The semi-assortative update (S8) correctly includes them, so this looks like an oversight rather than a design choice. Every omniassortative figure (Figs. 2–5, the relative gains, the disassortativity proportions) is produced by this algorithm. The authors need to correct the derivation and re-run before the empirical claims can be accepted.

A smaller flaw: the main text says full disassortativity requires negative γ, but the Supplementary construction (S11) zeroes the k>C diagonal and keeps rates nonnegative; negative γ would make the latent Poisson counts ill-defined. The paper should pick one formalization.

The remaining concerns are softer. The comparisons only pit the model against the strictly assortative baseline (Contisciani et al.), so the abstract's 'state-of-the-art' claim is overreach. And the identifiability guarantee depends on an imposed separability constraint; if real data lack pure communities, the theorem does not apply. That is a limitation to state, not a fatal flaw.

Who this is for: readers working on hypergraph community detection and tensor-based network models. The identifiability result and the synthetic generation algorithm are worth citing even if the current empirical results need re-verification.

Recommendation: send to a serious referee. The core model and theory are novel and plausible; the inference bug is real but fixable, and the authors ship code so a careful revision can verify the claims. This deserves a proper peer review, not a desk rejection.

Referee Report

3 major / 4 minor

Summary. The paper introduces Omni-Hype-SMT, a probabilistic model for hypergraphs in which nodes are softly assigned to latent classes and classes are softly assigned to latent communities. A low-rank CP-type factorization of the class-affinity tensor, with nonnegative parameters normalized to simplexes and a separability constraint on the class-community matrix, is claimed to be identifiable up to permutation. An EM algorithm is derived for three variants: strictly assortative, semi-assortative, and omniassortative. The empirical sections report improved heldout link prediction over a strictly assortative baseline, more interpretable node memberships, order-dependent disassortativity levels, and fast synthetic hypergraph generation. The paper also contains a fairly detailed supplementary note with derivations, identifiability theorems, and experimental details.

Significance. If the technical issues were resolved, this would be a useful contribution: it generalizes existing assortative hypergraph blockmodels, gives a concrete low-rank parameterization with a formal identifiability statement, and provides an algorithm for generating synthetic hypergraphs whose cost is linear in the expected number of hyperedges. The heldout link-prediction protocol is a genuine out-of-sample evaluation, and the case studies on drug interactions and political data illustrate potentially interesting structure. The main advertised strength—a tractable, identifiable model covering an assortative-to-disassortative spectrum—is exactly what is put in doubt by the inconsistencies described below, so the significance hinges on a fixable but load-bearing technical correction.

major comments (3)

[Supplementary Note 1, Eq. (S17) and Algorithm 2] The omniassortative E-step omits the community-order rates γ_k^(d). Under the latent Poisson construction in Eq. (S11), the multinomial thinning probabilities for the community counts A_ik^(d) conditional on A_i^(d) are proportional to γ_k^(d) times the bracket [∏_{i∈i} θ_i^T w_k − 1(k>C) Σ_c w_ck^d ∏_{i∈i} θ_ic], normalized by the sum of the same γ-weighted brackets. As written, Eq. (S17) and Algorithm 2 compute the expectation without any γ factor, so the algorithm is not computing the conditional expectation of the stated model. Consequently, the M-step updates in Eqs. (S12)–(S14) are not coordinate-wise maximizers of the evidence lower bound, and the EM convergence argument does not apply. Since every omniassortative experimental result—Figs. 2–5, the relative gains in Fig. 4, and the disassortativity proportions—is produced by this algorithm, the empirical claims currently rest on an unverified procedure.
[Section II B (Omniassortativity) vs. Supplementary Note 1 (Eq. S11)] There is an inconsistency between the main text's description and the implemented model. The main text states that a high degree of disassortativity requires some γ_k^(d) to be negative and refers to a careful scheme ensuring all elements of Λ remain nonnegative. However, Supplementary Note 1 implements a different construction: for k > C the diagonal elements are zeroed (Eq. S11) while γ_k^(d) are kept nonnegative. If γ_k^(d) were actually negative, the latent Poisson rates in Eq. (7) would be negative for many subcounts, making the generative model ill-defined. The paper should clarify which model is actually fit and how the negative-γ representation relates to the nonnegative zero-diagonal implementation; as written, the algorithm implements a model different from the one whose identifiability is advertised through Lemma A.8 and Corollary A.9.
[Section II B (Model identifiability) and Theorem A.5] The uniqueness theorem relies on the separability constraint W = [I_C | ...] in Eq. (5), which is imposed during fitting rather than verified from data. The statement that this constraint is 'satisfied in all of our experiments' means only that the algorithm enforced it; it does not provide evidence that real hypergraphs contain pure communities. If the true data-generating process lacks such pure communities, the identifiability guarantee does not apply, and the interpretation of the estimated classes and communities is not protected by the theorem. The paper should state this scope condition prominently and discuss its consequences, since the advertised 'ensuring model identifiability' is conditional on a structural assumption that is untestable from the observed hyperedges alone.

minor comments (4)

[Methods, Section V B (Proportion of disassortativity)] The disassortativity proportion is computed from the model's own latent allocations. This is a descriptive summary of the fitted model, not a validation of the model or of the discovered structure; the text should label it as such.
[Fig. 4b] The caption states that lower entropy H(Θ) denotes 'more interpretable class structure'; entropy measures mixed membership, and interpretability is an additional judgment. Consider rephrasing to avoid conflating the two.
[Supplementary Note 4, Theorem A.5 proof] The proof of Theorem A.5 states that the CP decomposition is unique 'up to permutation and scaling' and then uses the simplex constraints to remove scaling. The argument is plausible, but it would help the reader if the exact role of the l1 normalization of Θ and W were spelled out at the point where it is used to conclude ψ = 1.
[Table 1] The hospital dataset is described as 'semi-synthetic' in the main text but the table header does not indicate which rows are semi-synthetic; adding a footnote would improve clarity.

Circularity Check

1 steps flagged · score 2.0 of 10

Minor circularity: the synthetic-hypergraph validation uses DAWN-fitted parameters as both generator and reference; the central model derivation and heldout link-prediction comparisons are otherwise self-contained.

fitted input called prediction [Section II E ('Fast hypergraph generation') and Methods C ('Hypergraph generation'), Eq. (11)]
"To illustrate this algorithm empirically, we generate synthetic hypergraph data using the model parameters learned on the DAWN data, as described in Section II C. ... To check whether the synthetic data is similar to the true data, we compare a number of statistics: the node-degree distributions, hyperedge order distributions, inclusion occurrence distributions (39), and projected adjacency matrices. ... The plots appear nearly identical, demonstrating that the synthetic data closely resembles the true data."

The synthetic hypergraph is sampled from Omni-Hype-SMT with parameters fit to the DAWN data, and the 'true' reference data is that same DAWN data. The comparison therefore measures the model's in-sample reconstruction ability, not an independent demonstration that the generator produces realistic hypergraphs with prespecified structure. Both the generated and reference objects are functions of the same fitted parameters, so the apparent match is forced by construction rather than by an out-of-sample check.

full rationale

No significant circularity was found in the paper's central derivation. The identifiability claim is supported by a theorem proved in Supplementary Note 4 using external uniqueness results (Kruskal, Sidiropoulos-Bro, Hitchcock, Gillis), and the separability constraint W=[I_C | ...] is stated explicitly as an imposed assumption rather than hidden in the data. The claim that the strictly assortative model (Ref. 6) is a special case is verified by the equations themselves: setting W=I_C, K=C makes Eq. (3) diagonal and recovers the Hypergraph-MT model. The heldout link-prediction experiments are genuine out-of-sample evaluations, with masked hyperedges and multiple train-test splits, so those predictive gains are not forced by fitting. Self-citations to (6), (22), and (55) are not load-bearing: the relevant latent-subcount and thinning derivations are reproduced in Supplementary Note 1, and the cited O(1) computation trick is an implementation detail rather than a central premise. One minor circularity does exist in the synthetic-generation demonstration, where parameters learned on DAWN are used to generate a synthetic dataset and then compared back to the same DAWN data; this is an in-sample goodness-of-fit check, not an independent validation of the claimed generation capability. That step is peripheral to the model's main contributions, so the overall circularity score is low. The reviewer-flagged omission of gamma rates in Eq. (S17) is a potential correctness and convergence issue, but it is not an instance of circular reasoning because it does not reduce a prediction to its inputs by construction.

Assumptions & free parameters 4 free parameters · 6 assumptions · 2 invented entities

The model's central claims rest on the Poisson tensor model, the low-rank CP parameterization of class affinities, the separability constraint W=[I_C|...] for identifiability, and standard probabilistic and tensor uniqueness results. The fitted parameters Θ, W, and Γ are all data-driven, not derived from first principles, and the separability constraint is an imposed rather than learned premise.

free parameters (4)

node-class membership matrix Θ = estimated via EM (N×C)
Every node's soft assignment to latent classes; central to Eq. (2).
class-community membership matrix W = estimated via gradient ascent with δ=1e-6 (C×K)
Defines communities as mixtures of classes in Eq. (3) and carries the identifiability anchor W=[I_C|...].
community-order rates γ_k^(d) = estimated per order and community, initialized to 1
Rate of d-order interactions in community k; estimated by M-step Eqs. (S13)-(S14).
model dimensions C and K = chosen by grid search on heldout likelihood, e.g., DAWN C=16, K=48; see Table 1
The number of latent classes and communities is not determined by theory; it is selected by cross-validation.

assumptions (6)

domain assumption Conditional independence and Poisson counts for hyperedge entries (Eq. 1)
The likelihood treats all hyperedge counts as independent Poisson given rates; standard but not derived.
domain assumption Low-rank CP factorization of class affinity tensors, with classes interacting only within communities (Eq. 3)
This structural restriction makes estimation tractable but limits the representable family of affinity tensors.
ad hoc to paper Separability W=[I_C|...] and simplex constraints on Θ and W (Eq. 5)
Imposed to obtain identifiability; not derived from the data-generating process and may be false in real systems.
standard math Symmetric tensor CP uniqueness and NMF separability results (Refs. 28, 29, 65)
Used in Theorem A.5; assumes Kruskal rank conditions and full rank of Θ.
standard math Poisson additivity and multinomial thinning (Supplementary Note 2)
Basis for the latent subcount EM updates and for the hypergraph sampling algorithm.
ad hoc to paper For k>C, diagonal entries of non-pure communities are set to zero (Eq. S11)
This zero-diagonal constraint is what allows pure disassortativity without negative rate parameters; it is a modeling choice.

invented entities (2)

latent node classes independent evidence
purpose: Clusters of nodes with shared interaction profiles; the primary interpretable output, such as inferred drug classes.
Classes are validated indirectly by heldout link prediction and by post hoc agreement with known pharmacology, but there is no direct independent assay of the latent partition.
latent class-communities
purpose: Mixtures of classes that carry d-order interaction rates; they allow disassortative interactions between classes without a combinatorial affinity tensor.
Communities are not directly observed; the only external check is qualitative, such as the known 'Q-ball' combination, so independent evidence outside the fitted model is lacking.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Broad Spectrum Structure Discovery in Large-Scale Higher-Order Networks." pith.science (2026). https://pith.science/paper/33JHEWCU

@misc{pith2026250521748,
  author       = {Pith},
  title        = {Pith review of: Broad Spectrum Structure Discovery in Large-Scale Higher-Order Networks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/33JHEWCU}},
  note         = {Machine review of arXiv:2505.21748}
}

read the original abstract

Complex systems are often driven by higher-order interactions among multiple units, naturally represented as hypergraphs. Understanding dependency structures within these hypergraphs is crucial for understanding and predicting the behavior of complex systems but is made challenging by their combinatorial complexity and computational demands. In this paper, we introduce a class of probabilistic models that efficiently represents and discovers a broad spectrum of mesoscale structure in large-scale hypergraphs. The key insight enabling this approach is to treat classes of similar units as themselves nodes in a latent hypergraph. By modeling observed node interactions through latent interactions among classes using low-rank representations, our approach tractably captures rich structural patterns while ensuring model identifiability. This allows for direct interpretation of distinct node- and class-level structures. Empirically, our model improves link prediction over state-of-the-art methods and discovers interpretable structures in diverse real-world systems, including pharmacological and social networks, advancing the ability to incorporate large-scale higher-order data into the scientific process.

Figures

Figures reproduced from arXiv: 2505.21748 by the authors.

**Figure 2.** Omni-Hype-SMT recovers core-periphery structure among US Supreme Court [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Omni-Hype-SMT learns a latent hypergraph between identified drug classes in drug-drug [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗

Figures from the paper (2 more)

**Figure 4.** Figure 4: Relaxing strict assortativity improves interpretability and link prediction. a) For each dataset, we show relative gain in heldout log-likelihood over the strictly assortative baseline. Positive values indicate better link prediction for Omni-Hype-SMT, with error bars …

**Figure 5.** Figure 5: Generating synthetic data with Omni-Hype-SMT. [PITH_FULL_IMAGE:figures/full_fig_p009_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

67 extracted references · 67 canonical work pages

[1]

These results illustrate the importance of carefully modeling omniassortativity in the hypergraph setting to appropriately capture underlying latent structure. E. F ast hypergraph generation Finally, Omni-Hype-SMT is a generative model, and here we show how to use it to generate synthetic hyper- graphs with prespecified mesoscale structures. The prob- lem...

work page
[2]

Battiston, G

F. Battiston, G. Cencetti, I. Iacopini, V. Latora, M. Lu- cas, A. Patania, J.-G. Young, G. Petri, Networks beyond pairwise interactions: structure and dynamics. Physics Reports 874, 1–92 (2020)

work page 2020
[3]

M. A. Porter, J.-P. Onnela, P. J. Mucha, Communities in networks. Notices of the AMS 56 (2009)

work page 2009
[4]

Fortunato, Community detection in graphs

S. Fortunato, Community detection in graphs. Physics Reports 486, 75–174 (2010)

work page 2010
[5]

Antelmi, G

A. Antelmi, G. Cordasco, M. Polato, V. Scarano, C. Spagnuolo, D. Yang, A survey on hypergraph rep- resentation learning. ACM Computing Surveys 56, 1–38 (2023)

work page 2023
[6]

P. S. Chodrow, N. Veldt, A. R. Benson, Generative hy- pergraph clustering: From blockmodels to modularity. Science Advances 7, eabh1303 (2021)

work page 2021
[7]

Contisciani, F

M. Contisciani, F. Battiston, C. De Bacco, Inference of hyperedges and overlapping communities in hypergraphs. Nature Communications 13, 7229 (2022)

work page 2022
[8]

Ruggeri, M

N. Ruggeri, M. Contisciani, F. Battiston, C. De Bacco, Community detection in large hypergraphs. Science Ad- vances 9, eadg9159 (2023)

work page 2023

Show all 67 references

[9]

Papachristou, J

M. Papachristou, J. Kleinberg, Core-periphery models for hypergraphs, Proceedings of the 28th ACM SIGKDD Conference on Knowledge Discovery and Data Mining (2022), pp. 1337–1347

2022
[10]

Tudisco, D

F. Tudisco, D. J. Higham, Core-periphery detection in hypergraphs. SIAM Journal on Mathematics of Data Science 5, 1–21 (2023)

2023
[11]

Sales-Pardo, A

M. Sales-Pardo, A. Marin´ e-Tena, R. Guimer` a, Hyper- edge prediction and the statistical mechanisms of higher- order and lower-order interactions in complex networks. Proceedings of the National Academy of Sciences 120, e2303887120 (2023)

2023
[12]

Ruggeri, A

N. Ruggeri, A. Lonardi, C. De Bacco, Message-passing on hypergraphs: detectability, phase transitions and higher- order information. Journal of Statistical Mechanics: The- ory and Experiment 2024, 043403 (2024)

2024
[13]

Brusa, C

L. Brusa, C. Matias, Model-based clustering in simple hypergraphs through a stochastic blockmodel. Scandina- vian Journal of Statistics (2024)

2024
[14]

Veldt, A

N. Veldt, A. R. Benson, J. Kleinberg, Hypergraph cuts with general splitting functions. SIAM Review 64, 650– 685 (2022)

2022
[15]

Pister, M

A. Pister, M. Barthelemy, Stochastic block hypergraph model. Physical Review E 110, 034312 (2024)

2024
[16]

S.-e. Yoon, H. Song, K. Shin, Y. Yi, How much and when do we need higher-order information in hypergraphs? a case study on hyperedge prediction, Proceedings of The Web Conference 2020 (2020), pp. 2627–2633

2020
[17]

Purkait, T.-J

P. Purkait, T.-J. Chin, A. Sadri, D. Suter, Clustering with hypergraphs: the case for large hyperedges. IEEE Transactions on Pattern Analysis and Machine Intelli- gence 39, 1697–1711 (2016)

2016
[18]

Ruggeri, F

N. Ruggeri, F. Battiston, C. De Bacco, Framework to generate hypergraphs with community structure. Physi- cal Review E 109, 034309 (2024)

2024
[19]

Vanhems, A

P. Vanhems, A. Barrat, C. Cattuto, J.-F. Pinton, N. Khanafer, C. R´ egis, B.-a. Kim, B. Comte, N. Voirin, Estimating potential infection transmission routes in hos- pital wards using wearable proximity sensors. PLOS One 8, e73970 (2013)

2013
[20]

P. W. Holland, K. B. Laskey, S. Leinhardt, Stochastic blockmodels: First steps. Social Networks 5, 109–137 (1983)

1983
[21]

E. M. Airoldi, D. M. Blei, S. E. Fienberg, E. P. Xing, Mixed membership stochastic blockmodels. Journal of Machine Learning Research 9, 1981–2014 (2008)

2008
[22]

B. Ball, B. Karrer, M. E. Newman, Efficient and prin- cipled method for detecting communities in networks. Physical Review E 84, 036103 (2011)

2011
[23]

Schein, M

A. Schein, M. Zhou, D. M. Blei, H. Wallach, Bayesian Poisson Tucker decomposition for learning the structure of international relations, Proceedings of the 33rd Inter- national Conference on Machine Learning (2016), pp. 2810–2819

2016
[25]

Aguiar, D

I. Aguiar, D. Taylor, J. Ugander, A tensor factorization model of multilayer network interdependence (2023)

2023
[26]

J. Hood, A. J. Schein, The all0core tensor decomposi- tion for sparse count data, International Conference on Artificial Intelligence and Statistics (PMLR, 2024), pp. 4654–4662

2024
[27]

T. G. Kolda, B. W. Bader, Tensor Decompositions and Applications. SIAM Review 51, 455–500 (2009)

2009
[28]

L. R. Tucker, Some mathematical notes on three-mode factor analysis. Psychometrika 31, 279–311 (1966)

1966
[29]

F. L. Hitchcock, The expression of a tensor or a polyadic as a sum of products. Journal of Mathematics and Physics 6, 164–189 (1927)

1927
[30]

Gillis, Nonnegative Matrix Factorization (SIAM, 2020)

N. Gillis, Nonnegative Matrix Factorization (SIAM, 2020)

2020
[31]

A. R. Benson, R. Abebe, M. T. Schaub, A. Jadbabaie, J. Kleinberg, Simplicial closure and higher-order link pre- diction. Proceedings of the National Academy of Sciences 115, E11221–E11230 (2018)

2018
[32]

B. M. Waters, K. G. Joshi, Intravenous quetiapine- cocaine use (“q-ball”). American Journal of Psychiatry 164, 173–174 (2007)

2007
[33]

Stewart III, J

C. Stewart III, J. Woon, Congressional committee as- signments, 103rd to 114th congresses, 1993-2017: Senate. Cambridge, MA: Massachusetts Institute of Technology (2017)

2017
[34]

J. H. Fowler, Legislative cosponsorship networks in the us house and senate. Social Networks 28, 454–465 (2006)

2006
[35]

J. H. Fowler, Connecting the congress: A study of cosponsorship networks. Political Analysis 14, 456–487 (2006)

2006
[36]

G´ enois, C

M. G´ enois, C. L. Vestergaard, J. Fournet, A. Panisson, I. Bonmarin, A. Barrat, Data on face-to-face contacts in an office building suggest a low-cost vaccination strategy based on community linkers. Network Science 3, 326–347 (2015)

2015
[37]

Mastrandrea, J

R. Mastrandrea, J. Fournet, A. Barrat, Contact patterns in a high school: a comparison between data collected us- ing wearable sensors, contact diaries and friendship sur- veys. PLOS One 10, e0136497 (2015)

2015
[38]

Veldt, A

N. Veldt, A. R. Benson, J. Kleinberg, Combinatorial characterizations and impossibilities for higher-order ho- mophily. Science Advances 9, eabq3200 (2023)

2023
[39]

Kami´ nski, P

B. Kami´ nski, P. Pra lat, F. Th´ eberge, Hypergraph artifi- cial benchmark for community detection (h–abcd). Jour- nal of Complex Networks 11, cnad028 (2023). 12

2023
[40]

Q. F. Lotito, F. Musciotto, A. Montresor, F. Battiston, Higher-order motif analysis in hypergraphs. Communi- cations Physics 5, 79 (2022)

2022
[41]

M. S. Mariani, Z.-M. Ren, J. Bascompte, C. J. Tessone, Nestedness in complex networks: observation, emer- gence, and implications. Physics Reports 813, 1–90 (2019)

2019
[42]

N. W. Landry, J.-G. Young, N. Eikmeier, The simplicial- ity of higher-order networks. EPJ Data Science 13, 17 (2024)

2024
[43]

LaRock, R

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

2023
[44]

C. A. Joslyn, S. G. Aksoy, T. J. Callahan, L. E. Hunter, B. Jefferson, B. Praggastis, E. Purvine, I. J. Tripodi, Hypernetwork science: from multidimensional networks to computational topology, International Conference on Complex Systems (Springer, 2020), pp. 377–392

2020
[45]

M. Zhou, Y. Cong, B. Chen, Augmentable gamma belief networks. Journal of Machine Learning Research 17, 1– 44 (2016)

2016
[46]

Basbug, B

M. Basbug, B. Engelhardt, Hierarchical compound pois- son factorization, International Conference on Machine Learning (PMLR, 2016), pp. 1795–1803

2016
[47]

Gallo, L

L. Gallo, L. Lacasa, V. Latora, F. Battiston, Higher- order correlations reveal complex memory in temporal hypergraphs. Nature Communications 15, 4754 (2024)

2024
[48]

Iacopini, M

I. Iacopini, M. Karsai, A. Barrat, The temporal dynam- ics of group interactions in higher-order social networks. Nature Communications 15, 7391 (2024)

2024
[49]

Chowdhary, A

S. Chowdhary, A. Kumar, G. Cencetti, I. Iacopini, F. Battiston, Simplicial contagion in temporal higher- order networks. Journal of Physics: Complexity 2, 035019 (2021)

2021
[50]

X. He, P. S. Chodrow, P. J. Mucha, Hypergraph link prediction via hyperedge copying. arXiv preprint arXiv:2502.02386 (2025)

2025 arXiv
[51]

A. R. Benson, R. Kumar, A. Tomkins, Sequences of sets, Proceedings of the 24th ACM SIGKDD Interna- tional Conference on Knowledge Discovery & Data Min- ing (2018), pp. 1148–1157

2018
[52]

Badalyan, N

A. Badalyan, N. Ruggeri, C. De Bacco, Structure and inference in hypergraphs with node attributes. Nature Communications 15, 7073 (2024)

2024
[53]

Contisciani*, M

M. Contisciani*, M. Hobbhahn*, E. A. Power, P. Hennig, C. De Bacco, Flexible inference in heterogeneous and at- tributed multilayer networks. PNAS Nexus p. pgaf005 (2025)

2025
[54]

Gallo, G

G. Gallo, G. Longo, S. Pallottino, S. Nguyen, Directed hypergraphs and applications. Discrete Applied Mathe- matics 42, 177–201 (1993)

1993
[55]

A. P. Dempster, N. M. Laird, D. B. Rubin, Maxi- mum likelihood from incomplete data via the em algo- rithm. Journal of the Royal Statistical Society: Series B (Methodological) 39, 1–22 (1977)

1977
[56]

Schein, Allocative Poisson factorization for computa- tional social science (2019)

A. Schein, Allocative Poisson factorization for computa- tional social science (2019)

2019
[57]

Yıldırım, M

S. Yıldırım, M. B. Kurutmaz, M. Barsbey, U. S ¸im¸ sekli, A. T. Cemgil, Bayesian allocation model: marginal likelihood-based model selection for count tensors. IEEE Journal of Selected Topics in Signal Processing 15, 560– 573 (2020)

2020
[58]

Kullback, R

S. Kullback, R. A. Leibler, On information and suffi- ciency. The Annals of Mathematical Statistics 22, 79–86 (1951)

1951
[59]

Fawcett, An introduction to ROC analysis

T. Fawcett, An introduction to ROC analysis. Pattern recognition letters 27, 861–874 (2006)

2006
[60]

C. J. Wu, On the convergence properties of the em algo- rithm. The Annals of statistics pp. 95–103 (1983)

1983
[61]

A. G. Baydin, B. A. Pearlmutter, A. A. Radul, J. M. Siskind, Automatic differentiation in machine learning: a survey. Journal of machine learning research 18, 1–43 (2018)

2018
[62]

De Bacco, E

C. De Bacco, E. A. Power, D. B. Larremore, C. Moore, Community detection, link prediction, and layer interde- pendence in multilayer networks. Physical Review E 95, 042317 (2017)

2017
[63]

Contisciani, E

M. Contisciani, E. A. Power, C. De Bacco, Community detection with node attributes in multilayer networks. Scientific reports 10, 1–16 (2020)

2020
[64]

J. F. C. Kingman, Poisson processes, vol. 3 (Clarendon Press, 1992)

1992
[65]

Lin, Divergence measures based on the shannon en- tropy

J. Lin, Divergence measures based on the shannon en- tropy. IEEE Transactions on Information theory 37, 145–151 (1991)

1991
[66]

N. D. Sidiropoulos, R. Bro, On the uniqueness of multi- linear decomposition of n-way arrays. Journal of Chemo- metrics: A Journal of the Chemometrics Society 14, 229–239 (2000)

2000
[67]

Ben-Israel, T

A. Ben-Israel, T. N. Greville, Generalized inverses: the- ory and applications (Springer Science & Business Media, 2006)

2006
[68]

hyperevents

J. B. Kruskal, Rank, decomposition, and uniqueness for 3-way and n-way arrays. Multiway data analysis (1989), pp. 7–18. 13 V. METHODS A. Parameter estimation To estimate the parameters Θ , W, and Γ := γ (d) k d,k, given the input data A(:), we perform maximum like- lihood esti...

1989

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.

[1] [1]

These results illustrate the importance of carefully modeling omniassortativity in the hypergraph setting to appropriately capture underlying latent structure. E. F ast hypergraph generation Finally, Omni-Hype-SMT is a generative model, and here we show how to use it to generate synthetic hyper- graphs with prespecified mesoscale structures. The prob- lem...

work page

[2] [2]

Battiston, G

F. Battiston, G. Cencetti, I. Iacopini, V. Latora, M. Lu- cas, A. Patania, J.-G. Young, G. Petri, Networks beyond pairwise interactions: structure and dynamics. Physics Reports 874, 1–92 (2020)

work page 2020

[3] [3]

M. A. Porter, J.-P. Onnela, P. J. Mucha, Communities in networks. Notices of the AMS 56 (2009)

work page 2009

[4] [4]

Fortunato, Community detection in graphs

S. Fortunato, Community detection in graphs. Physics Reports 486, 75–174 (2010)

work page 2010

[5] [5]

Antelmi, G

A. Antelmi, G. Cordasco, M. Polato, V. Scarano, C. Spagnuolo, D. Yang, A survey on hypergraph rep- resentation learning. ACM Computing Surveys 56, 1–38 (2023)

work page 2023

[6] [6]

P. S. Chodrow, N. Veldt, A. R. Benson, Generative hy- pergraph clustering: From blockmodels to modularity. Science Advances 7, eabh1303 (2021)

work page 2021

[7] [7]

Contisciani, F

M. Contisciani, F. Battiston, C. De Bacco, Inference of hyperedges and overlapping communities in hypergraphs. Nature Communications 13, 7229 (2022)

work page 2022

[8] [8]

Ruggeri, M

N. Ruggeri, M. Contisciani, F. Battiston, C. De Bacco, Community detection in large hypergraphs. Science Ad- vances 9, eadg9159 (2023)

work page 2023