Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

HyperQuery: Beyond Binary Link Prediction

T0 review · 4 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read HyperQuery is a self-supervised message-passing framework that predicts missing hyperedges and their relation types, and the paper reports it outperforms prior hypergraph methods on several benchmarks by seeding node features with cluster…

desk verdict Useful combination of known pieces, but the SOTA numbers depend on an unstated detail—whether clustering sees test hyperedges—so the empirical claim is not yet supported. read the letter →

arxiv 2501.07731 v1 pith:CWUBUPP7 submitted 2025-01-13 cs.LG cs.SI

Sepideh Maleki , Josh Vekhter , Keshav Pingali This is my paper · ORCID
classification cs.LGcs.SI
keywords hyperedgepredictionknowledgehypergraphcompletionconvolutionmessagepassingself-supervisedlearningclusteringlinkbilinearpooling
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 sets out to show that one framework can answer both questions that hypergraph models raise: whether a given set of nodes forms a hyperedge, and, if so, what relation type that hyperedge carries. Its central move is to start a local message-passing convolution operator from global structure: every node is tagged with the id of the cluster it belongs to in a partition of the hypergraph, and those tags are pooled onto hyperedges as initial features. Training is self-supervised, using only hypergraph structure together with existing edge types. If the reported numbers hold, the framework improves on prior hypergraph methods by large margins on standard benchmarks, which would make cluster-based initialization a practical alternative to random-walk embeddings for higher-order link prediction.

What carries the argument

The Edge2Edge (E2E) convolution operator, defined by composing an edge-to-node step ($h^v_i = \text{AGG}(h^e, \forall e \in N(v_i))$) with a node-to-edge step ($h^S = \sigma(W^k \Omega(\{h^{k-1}_v \forall v \in S\}))$), where AGG is a harmonic-mean aggregation and Ω is a statistical summary (mean, variance, or minmax). The one-hot cluster id $x_v$ from the BiPart partition is concatenated to every node feature, and the final edge embedding is computed by flattening the outer product $\Omega^k_e \Omega^{k T}_e$ before a linear map. This cluster-id augmentation is what injects global hypergraph structure into the local message-passing updates.

What would settle it

Re-run the experiments with the cluster partition computed on the training hyperedges only (e.g., by holding out test hyperedges before running BiPart), keeping the same 20/10/70 split; compare MRR on Table 1 and AUC on Table 4. If the numbers fall substantially, the central claim of improvement is not supported; if they match, the claim survives this check.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single self-supervised message-passing framework, HyperQuery, solves both hyperedge prediction and knowledge hypergraph completion by learning hyperedge embeddings with an Edge2Edge convolution operator. The operator passes messages alternately from hyperedges to nodes (harmonic-mean aggregation) and from nodes back to hyperedges (a statistical summarizer Ω chosen as mean, variance, or minmax), then applies a learned weight matrix and, optionally, a bilinear outer-product flattening. The embedding is initialized using one-hot cluster ids obtained from a multilevel hypergraph partitioner (BiPart) for nodes and, for knowledge hypergraphs, from edge types. Training minimizes cross-entropy: reconstructing the one-hot type for knowledge hypergraphs, and separating true hyperedges from negative samples for hyperedge prediction. The paper reports that this recipe improves on HSimplE, HypE, NHP, HyperSAGNN, and HyperGCN across the FB-AUTO, M-FB15K, JF17K, iAF1260b, iJO1366, USPTO, and DBLP benchmarks.

Load-bearing premise

The reported gains rely on the clustering step being computed without using any test-hyperedge information; if BiPart partitions the full hypergraph including test hyperedges, the cluster ids would encode test answers and inflate the accuracy numbers.

Editorial extensions

If this is right

  • If the reported numbers hold, one framework with fixed hyperparameters (16 clusters, two convolution layers) covers both hyperedge prediction and knowledge hypergraph completion, removing the need for separate model families.
  • Cluster-id initialization could replace node2vec-style random walks as the default feature seed, avoiding random-walk sampling costs and the need to tune walk parameters.
  • The finding that mean aggregation works best for knowledge hypergraphs and minmax works best for simple hypergraphs gives practitioners a concrete rule for choosing the statistical summary Ω.
  • Since training needs no node attributes, the method applies directly to datasets such as protein complexes and chemical reactions where only hypergraph structure is available.

Reading between the lines

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

  • The paper leaves unspecified whether the clustering partition is computed on the training hyperedges only or on the full hypergraph; if BiPart sees test hyperedges, cluster ids could encode test co-occurrences and inflate the reported gains, so the partition-on-training-only rerun is the decisive check.
  • Because the paper uses a 20% train / 10% validation / 70% test split, repeating the evaluation with the conventional 70% train split would reveal whether the advantage persists when training data are abundant.
  • The same Edge2Edge operator could accept other global structural priors (spectral embeddings, motif counts, hierarchical partitions) in place of cluster ids, and the framework's performance on datasets where the cut objective is a poor fit is an open question.
  • If the leakage check is clean, HyperQuery's simplicity makes it a natural default baseline for future hypergraph prediction work.
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

4 major / 6 minor

Summary. The manuscript proposes HyperQuery, a message-passing hyperedge convolution framework for hyperedge prediction in simple hypergraphs and knowledge hypergraph completion. Initial node features are obtained from a deterministic hypergraph partitioner (BiPart) that assigns one-hot cluster ids, which are then concatenated with aggregated neighborhood embeddings. The framework uses edge-to-node and node-to-edge convolutions with a choice of aggregation functions (mean, variance, minmax) and optional bilinear pooling. The paper reports state-of-the-art MRR and AUC numbers on three knowledge hypergraph benchmarks (FB-AUTO, M-FB15K, JF17K) and four hyperedge prediction datasets (iAF1260b, iJO1366, USPTO, DBLP), plus hyperedge classification results on Cora, Citeseer, and PubMed.

Significance. If the reported numbers are trustworthy, HyperQuery offers a simple and effective alternative to random-walk-based feature initialization for hypergraph learning. The idea of using a deterministic partitioner to produce global structural features is interesting and the paper includes a useful ablation study and a time-complexity analysis. However, the empirical claims are substantially weakened by ambiguities in the experimental protocol, particularly the use of clustering as a preprocessing step without any statement that clustering is restricted to training edges. This concern is load-bearing because the initial node features (Equation 5) are one-hot cluster ids, and if test-edge information enters clustering, the reported gains would be inflated.

major comments (4)
  1. [Section 4.2, Table 4] The paper never states whether BiPart clustering is performed on the training hyperedges only or on the full hypergraph. Sections 4.1 and 4.2 describe clustering as a preprocessing step on 'the hypergraph', and the split for hyperedge prediction is 20% train, 10% validation, 70% test. If BiPart minimizes hyperedge cut over all hyperedges, then cluster assignments—and hence the one-hot partition ids x_v used in Equation (5)—encode co-occurrence of nodes in test hyperedges. This would leak test information into the features from which scores are predicted, inflating the MRR in Table 1 and AUC in Table 4. The authors must clarify the exact preprocessing protocol and, if clustering is currently run on the full hypergraph, rerun all experiments with clustering computed only on the training subgraph (or training+validation, with the partition then applied to test nodes without re-optimization). This is central to the SOTA claim.
  2. [Equation (2), Equation (5)] The paper states that HyperQuery uses a 20/10/70 train/validation/test split, while for the baselines it 'adopt[s] the same setting as the one used in the original papers'. The original NHP, HyperSAGNN, and HyperGCN papers typically use a 70/10/20 split. This asymmetry means HyperQuery is trained on much less data and tested on much more data than the baselines, so the comparisons in Table 4 are not apples-to-apples. If the baselines are evaluated under a different split, the improvements could be partly an artifact of the split rather than the method. The authors should either evaluate all methods under the same split or justify the reversed split and provide baseline results under that split.
  3. [Tables 1, 4, and 6] The edge-to-node convolution is defined via an aggregation function AGG in Equation (2), and the same symbol appears in Equation (5), but AGG is never specified. The text in Section 3.2 says AGG is a 'harmonic mean' but no formula or further description is given. This is essential for reproducibility: a reader cannot know whether AGG is a simple mean, a harmonic mean, a sum, or something else. The authors should define AGG explicitly in the text or as a separate equation.
  4. [Tables 1, 4, and 6] All experimental results are reported as single point estimates with no error bars, standard deviations, or significance tests. Given the very large claimed improvements (e.g., MRR 99.9 vs. 49.4 on JF17K in Table 1), it is important to know whether these numbers are stable across random initializations and data splits. The authors should report mean and standard deviation over at least 5 runs, or otherwise provide evidence of statistical significance.
minor comments (6)
  1. [Table 2] Table 2 uses the symbol |E| for what appears to be the number of entities (nodes), whereas the formal definition in Section 1 uses |E| for hyperedges. For instance, FB-AUTO is listed with |E|=3,410 but #train=6,778, so the column cannot represent the number of hyperedges. The header should be changed to |V| or 'entities' to avoid confusion.
  2. [Section 3.2] There are several grammatical and typographical errors, including incomplete sentences: 'in fact it can just as well be thought of as .' and 'as long as we take S = N (e) (neighborhood of hyperedge e) for a known hyperedge e.' The text should be revised for clarity.
  3. [Section 4.2, Negative sampling] The negative sampling procedure is described only qualitatively ('half of the vertices sampled from e the remaining half from V - e'). It is not stated how many negative examples are generated per positive hyperedge, whether negatives are constructed for validation/test sets, or how the AUC is computed over the test set. These details should be specified.
  4. [Appendix A] The ablation study on the number of clusters is mentioned but Figure 4 is not described in the text, and the actual curves are not visible in the manuscript. The authors should summarize the trend in the text (e.g., best at 16 clusters for most datasets) and make the figure self-contained.
  5. [Appendix A] Appendix A says 'In our experiments in Section 6' but the correct reference is Section 4. Also, the phrase 'membership' and other small typos appear throughout; a careful proofread is needed.
  6. [Introduction] The introduction contains 'emperical evaluations' (should be 'empirical'). There are also minor citation formatting issues (e.g., 'TranshH' in Section 4.1 should be 'TransH').

Circularity Check

1 steps flagged · score 6.0 of 10

SOTA claim reduces to cluster features that encode test hyperedge co-membership; paper never restricts BiPart clustering to training edges.

  1. fitted input called prediction [Section 4, 'Clustering algorithm' paragraph; Eq. (5); Section 4.2 data split]
    "In our framework, we use BiPart as a pre-processing step where we partition the nodes of a hypergraph first and the we use the one hot vector of their partition id as the initial feature of the nodes of the hypergraph. ... We used 70% of the hyperedges in these for test, 10% for validation and 20% for training."

    BiPart minimizes the hyperedge cut of Eq. (1) over the input hypergraph. The paper describes clustering as a preprocessing step on 'a hypergraph' and never restricts it to the 20% training hyperedges, so the one-hot x_v in Eq. (5) can encode co-membership of nodes in the 70% test hyperedges. The E2E convolution of Eq. (4) and the frozen linear classifier then score each test query from x_v of its nodes, making the predicted existence of a test hyperedge a function of the cluster assignment, which is itself a function of the same test hyperedge. The reported Table 1 MRR and Table 4 AUC gains therefore reduce, by construction of the reported preprocessing, to test-edge membership information baked into the initial features rather than to the hyperedge convolution alone.

full rationale

The paper's derivation chain is: BiPart clustering produces one-hot cluster ids x_v (Eq. 5), which seed the E2E convolution (Eq. 4), whose output feeds a frozen linear classifier to yield hyperedge scores. The only step that is circular is the unsupervised preprocessing: BiPart partitions nodes by minimizing cut over all hyperedges of the input hypergraph, and the Experimental section says 'we partition the nodes of a hypergraph first' before describing test splits of 70%, with no statement that clustering uses only training hyperedges. If the full hypergraph is used, x_v encodes whether nodes co-occur in test hyperedges, so the prediction for a test hyperedge is partly determined by that very hyperedge. This is a fitted-input-as-prediction pattern: the cluster assignment is fitted to the target edges and then used to 'predict' them. No separate circularity arises from the BiPart self-citation: BiPart is a published, independently evaluated tool, and the paper states 'any hypergraph partitioning algorithm can be used,' so the self-citation is not load-bearing. The convolution operator and its loss (Eq. 8) are standard supervised training, not circular. The HyperNetVec baseline in Appendix B is also only a baseline, not an argumentative premise. Overall, the central empirical claim is partially forced by the construction of the preprocessing step as written, but the method's equations themselves are not formally self-referential; hence score 6 rather than 8-10.

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

The method relies on a few hand-chosen hyperparameters (clusters, embedding size, layers, aggregation function) and a set of domain assumptions about clustering quality, negative sampling, and the integrity of the evaluation protocol. No new physical or conceptual entities are introduced.

free parameters (5)
  • Number of clusters = 16
    Chosen by hand; Figure 4 shows performance peaks at 16 for most datasets.
  • Embedding dimension = 64
    Fixed to 64 for all hyperedge prediction experiments, stated in Section 4.2.
  • Number of convolution layers = 2
    Stated in Section 4: 'We use 2 layers of hyperedge convolution in our test implementation'.
  • Aggregation function Omega = minmax for hyperedge prediction; mean for knowledge hypergraph completion
    Best variant selected per task from Table 4 and Table 1; all variants are reported.
  • Bilinear pooling enabled = yes for final model
    Ablation in Table 4 shows it helps on larger datasets; the final HQ model uses it.
assumptions (5)
  • domain assumption Hyperedge cut minimization produces node clusters that are semantically meaningful for prediction.
    Section 3.1 states the insight but gives no proof; the connection between clusters and predictive features is assumed.
  • domain assumption Neighborhood aggregation of hyperedge embeddings into node embeddings is a useful summary of local structure.
    Equation 2 defines E2N as an aggregation AGG, but the specific form (described as harmonic mean) is never specified, so the usefulness is assumed.
  • domain assumption Negative hyperedges sampled by mixing half of a true hyperedge with half of random nodes are valid non-edges.
    Section 3.5 and 4.2 adopt this strategy from NHP, motivated by chemical reactions, but it may not reflect all domains.
  • domain assumption Clustering is performed without using test-set hyperedge information.
    The paper does not state whether BiPart partitions the full hypergraph or only the training subgraph; if the former, the evaluation is leaked.
  • domain assumption The hypergraph representation is a complete and correct model of the underlying n-ary relations.
    The task definitions treat hyperedges as ground truth, which is standard but is an assumption about data quality.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperQuery: Beyond Binary Link Prediction." pith.science (2026). https://pith.science/paper/CWUBUPP7

@misc{pith2026250107731,
  author       = {Pith},
  title        = {Pith review of: HyperQuery: Beyond Binary Link Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/CWUBUPP7}},
  note         = {Machine review of arXiv:2501.07731}
}
read the original abstract

Groups with complex set intersection relations are a natural way to model a wide array of data, from the formation of social groups to the complex protein interactions which form the basis of biological life. One approach to representing such higher order relationships is as a hypergraph. However, efforts to apply machine learning techniques to hypergraph structured datasets have been limited thus far. In this paper, we address the problem of link prediction in knowledge hypergraphs as well as simple hypergraphs and develop a novel, simple, and effective optimization architecture that addresses both tasks. Additionally, we introduce a novel feature extraction technique using node level clustering and we show how integrating data from node-level labels can improve system performance. Our self-supervised approach achieves significant improvement over state of the art baselines on several hyperedge prediction and knowledge hypergraph completion benchmarks.

Figures

Figures reproduced from arXiv: 2501.07731 by the authors.

Figure 1
Figure 1. HyperQuery processing pipeline: Consider the hypergraph structured data illustrated above (left). Our system predicts the label of the red hyperedge by performing regression (right) based on an embedding (center, grey line) produced by the hyperedge convolution operation that we developed. hypergraph KH = (E, R) and a tuple of entities (e1, e2, ..., ek), we want to predict if this tuple forms a hyperedge and if so, … view at source ↗
Figure 2
Figure 2. Generating Useful Labels using clustering: Not all hypergraph data comes with labels (left). We leverage hypergraph clustering to first assign cluster id as labels to the nodes, obtaining an initial cluster assignment h 0 v (center), which we then propagate to the hyperedges via Max Pooling (right) in order to obtain the initial edge cluster assignment h 0 e. 3. Self-supervised Higher-order Link Prediction We begin … view at source ↗
Figure 3
Figure 3. An Edge2Edge Convolution Operator: Here we illustrate the function [h k v, hk e ] = E2E(S, hk−1 e ). Our query set S is annotated in blue on (left). For each vi ∈ S, we perform an E2N aggregation, and in the process, compute h k v (center left). We then summarize the distribution of node features using some choice of function Ω to obtain a statistical feature we call hsk e (center right). Passing these features thro… view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Performance of HyperQuery for different numbers of clusters. Next, we study the effect of the number of clusters in our dataset [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]
Figure 5
Figure 5. Figure 5: Learning W1 . Given a hypergraph with labels on nodes and edges, we experiment with a zoo of choices for Ω. In this illustration, we convolve once, multiply by W1 and σ and then pass through a fully connected layer to return to a space of the same dimension of hyperedg…

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. Two-dimensional Taxonomy for N-ary Knowledge Representation Learning Methods

    cs.LG 2025-06 conditional novelty 6.0 of 10

    A survey of n-ary knowledge representation learning methods proposes a two-dimensional taxonomy based on modeling technique and entity role/position awareness.

Reference graph

Works this paper leans on

19 extracted references · 8 canonical work pages · cited by 1 Pith paper

  1. [2]

    URL https://hal.archives-ouvertes.fr/hal-02999888

    doi: 10.1145/3336191.3371800. URL https://hal.archives-ouvertes.fr/hal-02999888. Devine, K. D., Boman, E. G., Heaphy, R. T., Bisseling, R. H., and Catalyurek, U. V . Parallel hypergraph partitioning for scientific computing. In Proceedings of the 20th International Conference on Parallel and Distributed Processing , IPDPS’06, pp. 124–124, Washington, DC, USA,

    arXiv

  2. [5]

    Hypergraph Models of Biological Networks to Identify Genes Critical to Pathogenic Viral Response

    URL https://arxiv.org/abs/2010.03068. Fiduccia, C. and Mattheyses, R. A linear-time heuristic for improving network partitions. In 19th Design Automation Conference, pp. 175–181,

    work page Pith review arXiv 2010

  3. [13]

    Wen, J., Li, J., Mao, Y ., Chen, S., and Zhang, R

    URL https://arxiv.org/abs/2105.08476. Wen, J., Li, J., Mao, Y ., Chen, S., and Zhang, R. On the representation and embedding of knowledge bases beyond binary relations,

    arXiv

  4. [15]

    ISBN 9781450368599

    Association for Computing Machinery. ISBN 9781450368599. doi: 10.1145/3340531.3411870. URL https://doi.org/10.1145/3340531.3411870. Yang, B., Yih, W.-t., He, X., Gao, J., and Deng, L. Embedding entities and relations for learning and inference in knowledge bases,

    arXiv

  5. [16]

    Zhang, R., Zou, Y ., and Ma, J

    URL https://arxiv.org/abs/1412.6575. Zhang, R., Zou, Y ., and Ma, J. Hyper-SAGNN: a self-attention based graph neural network for hypergraphs. InInternational Conference on Learning Representations (ICLR) ,

    arXiv

  6. [17]

    URL https://proceedings.neurips.cc/paper/2006/file/dff8e9c2ac33381546d96deea99 22999-Paper.pdf. Zhu, H. and Koniusz, P. Graph convolutional network with generalized factorized bilinear aggregation,

    work page 2006

  7. [18]

    Graph Convolutional Network with Generalized Factorized Bilinear Aggregation

    URL https://arxiv.org/abs/2107.11666. Zitnik, M., Agrawal, M., and Leskovec, J. Modeling polypharmacy side effects with graph convolutional networks. Bioinformatics, 34(13):457–466,

    work page Pith review arXiv

  8. [19]

    We used 20% of the hyperedges in these datasets for test, 10% for validation and 70% for training (for 20 epochs). Since there are no previous approaches for hyperedge classification, we modified existing methods as baselines for our problem: HyperNetVec (Maleki et al., 2022): an unsupervised multi-level approach to generate the representation of a hyperg...

    work page 2022

Show all 19 references
  1. [1927]

    URL https://onlinelibrary.wiley.com/do i/abs/10.1002/sapm192761164

    doi: https://doi.org/10.1002/sapm192761164. URL https://onlinelibrary.wiley.com/do i/abs/10.1002/sapm192761164. Karypis, G., Aggarwal, R., Kumar, V ., and Shekhar, S. Multilevel hypergraph partitioning: Applications in vlsi domain. IEEE Trans. V ery Large Scale Integr . Syst.,...

    doi

  2. [1982]

    Giurgiu, M., Reinhard, J., Brauner, B., Dunger-Kaltenbach, I., Fobo, G., Frishman, G., Montrone, C., and Ruepp, A

    doi: 10.1109/DAC.1982.1585498. Giurgiu, M., Reinhard, J., Brauner, B., Dunger-Kaltenbach, I., Fobo, G., Frishman, G., Montrone, C., and Ruepp, A. CORUM: the comprehensive resource of mammalian protein complexes-2019. Nucleic Acids Research,

    1982

  3. [2006]

    ISBN 1-4244-0054-6

    IEEE Computer Society. ISBN 1-4244-0054-6. URL http: //dl.acm.org/citation.cfm?id=1898953.1899056. Deylami, H. A. and Asadpour, M. Link prediction in social networks using hierarchical community detection. In 2015 7th Conference on Information and Knowledge Technology (IKT) , pp. 1–5,

    2015

  4. [2007]

    Tu, K., Cui, P., Wang, X., Wang, F., and Zhu, W

    doi: 10.1103/physreve.76.036106. Tu, K., Cui, P., Wang, X., Wang, F., and Zhu, W. Structural deep embedding for hyper-networks,

    doi

  5. [2014]

    ISBN 978-1-4503-2956-9

    ACM. ISBN 978-1-4503-2956-9. doi: 10.1145/2623330.2623732. URL http: //doi.acm.org/10.1145/2623330.2623732. Pi˜nero, J., Ram´ırez-Anguita, J. M., Sa¨uch-Pitarch, J., Ronzano, F., Centeno, E., Sanz, F., and Furlong, L. I. The DisGeNET knowledge platform for disease genomics: 20...

  6. [2015]

    HyperQuery: Beyond Binary Link Prediction Fatemi, B., Taslakian, P., Vazquez, D., and Poole, D

    doi: 10.1109/IKT.2015.7288742. HyperQuery: Beyond Binary Link Prediction Fatemi, B., Taslakian, P., Vazquez, D., and Poole, D. Knowledge hypergraphs: Prediction beyond binary relations. IJCAI’20,

    2015

  7. [2016]

    Yadati, N., Nimishakavi, M., Yadav, P., Nitin, V ., Louis, A., and Talukdar, P

    URL https://arxiv.org/abs/1604.08642. Yadati, N., Nimishakavi, M., Yadav, P., Nitin, V ., Louis, A., and Talukdar, P. Hypergcn: A new method of training graph convolutional networks on hypergraphs,

    arXiv

  8. [2017]

    Langley, P

    doi: 10.1109/TPAMI.2016.2545667. Langley, P. Crafting papers on machine learning. In Langley, P. (ed.), Proceedings of the 17th International Conference on Machine Learning (ICML

    2016

  9. [2019]

    doi: 10.1093/nar/gkz1021

    ISSN 0305-1048. doi: 10.1093/nar/gkz1021. URL https://doi.org/10.1093/nar/gkz1021. Raghavan, U. N., Albert, R., and Kumara, S. Near linear time algorithm to detect community structures in large-scale networks. Physical Review E, 76(3), sep

    doi

  10. [2020]

    doi: https: //doi.org/10.1016/j.physrep.2020.05.004

    ISSN 0370-1573. doi: https: //doi.org/10.1016/j.physrep.2020.05.004. URL https://www.sciencedirect.com/science/article/pi i/S0370157320302489. Networks beyond pairwise interactions: Structure and dynamics. Bhowmick, A. K., Meneni, K., Danisch, M., Guillaume, J.-L., and Mitra, ...

    2020 doi

  11. [2021]

    Maleki, S., Saless, D., Wall, D

    doi: 10.1145/3437801.3441611. Maleki, S., Saless, D., Wall, D. P., and Pingali, K. Hypernetvec: Fast and scalable hierarchical embedding for hypergraphs. In Ribeiro, P., Silva, F., Mendes, J. F., and Laureano, R. (eds.),Network Science, pp. 169–183, Cham,

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.