Pith. sign in

REVIEW 4 major objections 6 minor 29 references

Enhancing the Utility of Higher-Order Information in Relational Learning

T0 review · 4 major / 6 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 Graph-level architectures applied to hypergraph expansions outperform hypergraph-level ones, and hypergraph-level encodings provide the largest gains when paired with graph-level models.

desk verdict The hypergraph encodings are a genuine contribution; the headline claim that graph-level models beat hypergraph-level ones is not supported by the experiments as run. read the letter →

arxiv 2502.09570 v1 pith:S7SEWF7H submitted 2025-02-13 cs.LG stat.ML

Raphael Pellegrin , Lukas Fesser , Melanie Weber This is my paper
classification cs.LGstat.ML
keywords higher-orderinformationhypergraphneuralnetworksgraphencodingspositionalstructuralexpressivityWeisfeiler-Lehmantest
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to settle how best to exploit higher-order (multi-way) relations in relational learning. It argues that, despite the popularity of dedicated hypergraph neural networks, standard graph neural networks applied to the hypergraph's clique expansion often match or beat them, even on data natively stored as hypergraphs. It then proposes hypergraph-level encodings, features computed from hypergraph Laplacians, random walks, curvatures, and degree profiles, and shows empirically that these give large accuracy gains when added to graph-level models, while barely helping hypergraph-level ones. The theoretical part proves that these encodings make message-passing GNNs strictly more expressive than the 1-WL test and than their graph-level-encoding counterparts. If the results hold, the practical recipe is to prefer a standard GNN augmented with hypergraph encodings over a dedicated hypergraph message-passing network.

What carries the argument

The load-bearing objects are hypergraph-level encodings. H-k-LAPE uses the top k eigenvectors of the 0-Hodge Laplacian $H_0 = B_1^{\mathsf{T}} B_1$; H-k-RWPE records return probabilities of equal-nodes and equal-edges random walks on the hypergraph; HCP summarizes the Forman or Ollivier curvature of all hyperedges containing a node; H-LDP summarizes the degree distribution over a node's hypergraph neighborhood. Each encoding is computed as a preprocessing step and concatenated to the node features of the clique-expanded graph. The theoretical engine is a set of expressivity arguments: for each encoding, the paper exhibits pairs of non-isomorphic, 1-WL-indistinguishable graphs from the BREC dataset that the graph-level analogue of the encoding cannot separate, proving that the hypergraph-level version adds representational power that message passing alone lacks.

What would settle it

Retrain GCN and UniGCN on the five hypergraph node-classification datasets (Cora-CA, Cora-CC, Citeseer, PubMed, DBLP) with identical hidden dimension, learning rate, dropout, number of layers, and epochs, then compare test accuracy; if UniGCN matches or exceeds GCN under equal budgets, the paper's headline comparison is not supported.

Watch

Extended reading notes

Core claim

The paper's central claim is that hypergraph-level architectures do not, in general, deliver the expected benefit from higher-order structure: on five hypergraph node-classification datasets (Cora-CA, Cora-CC, Citeseer, PubMed, DBLP), GCN without encodings outperforms UniGCN, and the appendix reports the same for UniGIN, UniGAT, UniSAGE, and UniGCNII. The paper's proposed alternative is to keep the graph-level model but inject higher-order information as encodings computed on the hypergraph: Hodge-Laplacian eigenvectors (H-k-LAPE), hypergraph random-walk positional encodings (H-k-RWPE), hypergraph curvature profiles (HCP), and hypergraph local degree profiles (H-LDP). These encodings, the paper shows, capture structural information that 1-WL message passing cannot, and they produce substantial performance gains on graph classification benchmarks, especially social networks (Collab, Imdb, Reddit), when combined with GCN or GraphGPS, while they barely change the performance of hypergraph-level models. Theorems 3.2, 3.5, 3.11, and 3.13 establish that each proposed encoding strictly increases the expressivity of message-passing GNNs beyond 1-WL and beyond the corresponding graph-level encoding.

Load-bearing premise

The central comparison assumes that the default hyperparameters used for the hypergraph message-passing networks (hidden dimension 8, learning rate 0.01, 200 epochs, dropout 0.6) are a fair match for the graph-level networks (hidden dimension 128, learning rate 0.001, 300 epochs, dropout 0.2); if the capacity and budget gap explains the lower accuracy of UniGCN, the claim that graph-level architectures outperform hypergraph-level ones loses its empirical support.

Editorial extensions

If this is right

  • A default recipe for hypergraph node classification becomes: take the clique expansion, compute a hypergraph-level encoding such as HCP-FRC or H-k-RWPE, and train a standard GCN.
  • On graph classification benchmarks, hypergraph-level encodings outperform their graph-level analogues for GCN and GraphGPS, with the largest gains on social-network datasets (Collab, Imdb, Reddit) and on Mutag.
  • The expressivity theorems imply that these encodings let a message-passing GNN distinguish some non-isomorphic graphs that 1-WL, LAPE, LDP, and graph-level Forman curvature cannot, which is direct theoretical support for the empirical gains.
  • GIN, which is already 1-WL-maximal, also improves with hypergraph encodings, indicating that the encodings contribute information beyond Weisfeiler-Lehman expressivity alone.

Reading between the lines

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

  • If the GCN-versus-UniGCN gap disappears when hyperparameters are matched, the paper's practical recommendation would shift from 'avoid hypergraph message-passing' to 'hypergraph message-passing needs better default tuning'; the encodings and expressivity results would be unaffected.
  • The complexity estimates reported for the encodings (Lanczos H-LAPE in O(|E|k), H-RWPE in O(|V|d_max^k), HCP-FRC in O(|E| e_max)) suggest they are cheap enough to apply to large hypergraphs, though the paper does not run large-scale scalability tests.
  • Because the authors lift ordinary graphs to hypergraphs to compute encodings, any graph dataset can in principle be augmented with hypergraph-level descriptors, and the gains on social networks suggest this is most valuable where multi-way interactions are natural.
  • The lack of benchmarks with genuinely irreducible hypergraph structure, which the authors acknowledge, means the observed advantage of graph-level models may underestimate the value of hypergraph-level architectures on truly multi-way data.
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. This paper compares graph-level message-passing architectures (GCN, GIN, GPS) applied to hypergraph clique expansions with hypergraph-level architectures (UniGCN and related models), and proposes four families of hypergraph-level encodings: Hodge/normalized Laplacian positional encodings (H-LAPE), random-walk positional encodings (H-RWPE), hypergraph curvature profiles (HCP), and hypergraph local degree profiles (H-LDP). The central empirical claims are that graph-level architectures often outperform hypergraph-level ones even on natively hypergraph data, that the proposed encodings substantially improve graph-level models while not helping hypergraph-level models, and that the encodings provably increase MPGNN expressivity beyond 1-WL. Experiments span five hypergraph node-classification datasets and several graph-classification benchmarks, with expressivity checks on the BREC dataset and Rook/Shrikhande graphs. The paper also provides code and detailed appendices.

Significance. If the headline empirical claims are established under controlled conditions, the paper offers a practically useful and falsifiable recommendation: a standard GNN augmented with hypergraph encodings can be a better choice than a dedicated hypergraph message-passing network for many tasks. The proposed encoding family is a useful contribution, and the negative result that hypergraph encodings do not help the evaluated HNNs is a valuable benchmark observation. The authors also ship code and ground their expressivity statements in concrete, independent witness examples (BREC pairs, Rook and Shrikhande graphs). However, the main GCN-versus-UniGCN comparison is currently confounded by unequal hyperparameter budgets, and the universal claim that hypergraph encodings are always at least as beneficial as graph-level analogues is contradicted by Table 3. These issues affect load-bearing conclusions, so the paper requires major revision.

major comments (4)
  1. [§4.1, Table 7, Appendix F.1] The statement in §4.1 that "settings and optimization hyperparameters are held constant across tasks and baseline models" is contradicted by the appendices: GCN in Table 2 uses 3 layers, hidden dimension 128, learning rate 0.001, dropout 0.2, and 300 epochs (Table 7), while UniGCN in Table 2 uses 2 layers, hidden dimension 8, learning rate 0.01, dropout 0.6, and 200 epochs (Appendix F.1). With these unequal budgets, the accuracy gap in Table 2 (e.g., 75.66 vs. 71.05 on DBLP) cannot be attributed to graph-level versus hypergraph-level message passing. A matched-capacity comparison or a sensitivity analysis is required before the headline outperformance claim is justified.
  2. [§4.3, Table 3] The claim that "encodings computed at the hypergraph level are always at least as beneficial as their cousins computed at the graph level" is contradicted by Table 3: on Enzymes, GCN with LCP-ORC reaches 33.17 ± 1.43 while GCN with HCP-ORC reaches 32.83 ± 1.36, and on Peptides-s the graph-level LCP-ORC MAE is 0.252 versus 0.258 for HCP-ORC. The universal statement should be replaced by a qualified, per-dataset claim.
  3. [Theorems 3.2, 3.5, 3.11, 3.13] The proofs of Theorems 3.2, 3.5, 3.11, and 3.13 are witness-based: they exhibit one or two specific graph pairs whose encodings differ according to computations in Appendix E. This is sufficient for an existence proof of strict expressivity, but the theorem statements are written as general results and should explicitly state that they are proved by constructing a separating example. The paper should also document the numerical precision, eigenvector sign handling, and isomorphism verification used in those computations. Moreover, Table 12 shows that on several BREC categories the hypergraph-level encoding distinguishes fewer pairs than its graph-level counterpart (e.g., Hodge H-1-LAPE 98% vs. Hodge 1-LAPE 100% on Regular), so the relation should be described as an existence-based strict expressivity gain, not a dominance relation.
  4. [§4.2, Appendix F.1] The conclusion that hypergraph-level encodings do not improve hypergraph architectures is based on HNNs trained with only 2 layers, 8 hidden units, and 200 epochs (Appendix F.1), while GCN receives a much larger capacity and training budget (Table 7). The null result for UniGCN, UniGIN, UniGAT, UniSAGE, and UniGCNII may reflect undertrained baselines rather than a property of the encodings. A matched-budget hypergraph-level comparison, or at least an explicit discussion of this confound, is needed to support the paper's negative claim.
minor comments (6)
  1. [Definition 3.9] In Eq. (3.6), d(s) is written as a set, but it should be the scalar maximum distance; additionally, the formula for a hyperedge e in Eq. (3.7) drops the denominator d(e) without explanation, so the relation between the two formulas should be clarified.
  2. [Appendix E.1] For the Shrikhande graph's lifting, the edge degree matrix is stated as De = 3I8; the graph has 16 vertices of degree 6 and 32 triangular hyperedges, so the matrix should be 3I32. Please correct or explain the notation.
  3. [Tables 2 and 3] Several rows label the random-walk encoding as "H-19-R WPEE" (Tables 2 and 3), which appears to be a typo for "H-19-R WPE"; please standardize the encoding names.
  4. [§4.3] The paper says that hypergraph encodings "significantly boost performance on almost all datasets (only Proteins is not statistically significant)", but no statistical test is described; please state the test procedure and report the resulting p-values or confidence intervals.
  5. [Introduction vs. §1.2] The introduction says graph-level architectures "strictly outperform current hypergraph-level ones", while the contribution list says "comparable or superior performance"; align the wording to avoid overclaiming.
  6. [Table 12] The column header "Hyperaph (LDP)" contains a typo and should read "Hypergraph (H-LDP)".

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: encodings are fixed hypergraph invariants, expressivity proofs use external benchmark pairs, and performance claims are measured on fixed benchmarks.

full rationale

The derivation chain is self-contained. The proposed encodings (H-k-LAPE, H-k-RWPE, HCP, H-LDP) are defined as deterministic functions of classical hypergraph characteristics—Hodge/normalized Laplacian eigenvectors, equal-nodes/equal-edges random-walk return probabilities, and Forman/Ollivier curvature or degree profiles—so they are not fitted to the benchmark accuracies and no fitted parameter is later reported as a prediction. The expressivity theorems (3.2, 3.5, 3.11, 3.13, A.1) are justified by explicit distinguishing computations on external examples: BREC 'Basic' pair 0 and the Rook/Shrikhande srg(16,6,2,2) pair, with the per-encoding values tabulated in Appendix E; these examples do not presuppose the conclusion that the encodings are more expressive. The main empirical comparison (Table 2) is a fixed-benchmark evaluation: GCN and UniGCN accuracies are measured on standard splits rather than derived from any fitted parameter of the proposed method. The unequal training settings between GCN (Table 7) and UniGCN (Appendix F.1) are a legitimate fairness and validity concern for the claim that graph-level architectures outperform hypergraph-level ones, but that is an experimental confound, not a circular reduction: no equation in the paper makes the outcome equal to its input. Self-citations such as Fesser and Weber (2024a) are used only to define the graph-level LCP baselines and as context; the novel hypergraph-level encodings and their expressivity statements are verified with independent computations. Hence no step reduces the paper's claims to its own inputs.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claims rest on the lifting operation, standard WL equivalence, and the specific benchmark examples. The only tuned quantities are the encoding dimensions and the choice of aggregation for ORC. No new particles, forces, or physical entities are introduced.

free parameters (3)
  • Encoding dimension k for H-LAPE = 20 (H-20-LAPE)
    The number of Hodge or normalized Laplacian eigenvectors used as positional encodings is tuned on validation data, as stated in Section 4.1.
  • Encoding dimension k for H-RWPE = 19 (H-19-RWPE)
    The number of random walk steps used for hypergraph random-walk positional encodings is tuned on validation data, as stated in Section 4.1.
  • Aggregation function in H-ORC = average of pairwise Wasserstein distances
    The definition of H-ORC in Definition 3.9 leaves AGG(·) unspecified and the paper selects averaging; other aggregation choices would change the encoding.
assumptions (4)
  • domain assumption The lifting of a graph to a hypergraph is defined by adding hyperedges to groups of nodes that are pairwise interconnected.
    The expressivity proofs and empirical experiments depend on this specific lifting, but Appendix A.2 does not specify whether all cliques or only maximal cliques are used.
  • standard math MPGNNs are exactly as expressive as the 1-WL test.
    This equivalence, due to Xu et al. (2018), grounds the expressivity framework used in Section 2.3 and throughout the theorems.
  • domain assumption The BREC dataset pairs and the Rook and Shrikhande graphs have the stated isomorphism, spectral, and curvature properties.
    The proofs of Theorems 3.2, 3.5, 3.11, 3.13, and A.1 rely on these external facts as computed in Appendix E.
  • domain assumption Clique expansion is the best expansion for graph-level models on hypergraph data.
    Section 4.2 states this was found empirically, but no systematic comparison of expansion choices is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing the Utility of Higher-Order Information in Relational Learning." pith.science (2026). https://pith.science/paper/S7SEWF7H

@misc{pith2026250209570,
  author       = {Pith},
  title        = {Pith review of: Enhancing the Utility of Higher-Order Information in Relational Learning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/S7SEWF7H}},
  note         = {Machine review of arXiv:2502.09570}
}
read the original abstract

Higher-order information is crucial for relational learning in many domains where relationships extend beyond pairwise interactions. Hypergraphs provide a natural framework for modeling such relationships, which has motivated recent extensions of graph neural network architectures to hypergraphs. However, comparisons between hypergraph architectures and standard graph-level models remain limited. In this work, we systematically evaluate a selection of hypergraph-level and graph-level architectures, to determine their effectiveness in leveraging higher-order information in relational learning. Our results show that graph-level architectures applied to hypergraph expansions often outperform hypergraph-level ones, even on inputs that are naturally parametrized as hypergraphs. As an alternative approach for leveraging higher-order information, we propose hypergraph-level encodings based on classical hypergraph characteristics. While these encodings do not significantly improve hypergraph architectures, they yield substantial performance gains when combined with graph-level models. Our theoretical analysis shows that hypergraph-level encodings provably increase the representational power of message-passing graph neural networks beyond that of their graph-level counterparts.

Figures

Figures reproduced from arXiv: 2502.09570 by the authors.

Figure 1
Figure 1. A pair of graph from the BREC ”Basic” category (top left), the graphs’ liftings (top [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Example of a clique expansion of a hypergraph to a graph. The plots are created [PITH_FULL_IMAGE:figures/full_fig_p019_2.png] view at source ↗
Figure 3
Figure 3. Lifting of a graph to a hypergraph. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_3.png] view at source ↗
Figures from the paper (8 more)
Figure 4
Figure 4. Figure 4: Histogram of the number of nodes in the graphs in BREC. The bars are stacked. We [PITH_FULL_IMAGE:figures/full_fig_p027_4.png]
Figure 5
Figure 5. Figure 5: Histogram of the number of edges in the graphs in BREC. The bars are stacked. We [PITH_FULL_IMAGE:figures/full_fig_p027_5.png]
Figure 6
Figure 6. Figure 6: The Rook graph and the Shrikhande graph. Those two non-isomorphic graphs can be [PITH_FULL_IMAGE:figures/full_fig_p029_6.png]
Figure 7
Figure 7. Figure 7: The Rook graph (left), its lifting (right) and its lifting’s bipartite representation [PITH_FULL_IMAGE:figures/full_fig_p030_7.png]
Figure 8
Figure 8. Figure 8: The pair 0 of the ”Basic” category in BREC. Top: the two graphs in the pair. Second [PITH_FULL_IMAGE:figures/full_fig_p032_8.png]
Figure 9
Figure 9. Figure 9: The pair 0 of the ”Basic” category in BREC. Top: the two graphs in the pair. Second [PITH_FULL_IMAGE:figures/full_fig_p033_9.png]
Figure 10
Figure 10. Figure 10: The pair 0 of the regular category in BREC. Top: the two graphs in the pair. Second [PITH_FULL_IMAGE:figures/full_fig_p037_10.png]
Figure 11
Figure 11. Figure 11: The pair 0 of the strongly regular category in BREC. Top: the two graphs in the [PITH_FULL_IMAGE:figures/full_fig_p038_11.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

29 extracted references · 19 canonical work pages

  1. [1]

    A.4 Laplacians Several notions of Laplacians have been studied on hypergraphs. In this work, we consider two types of Laplacians on graphs for implementing H-LAPE, the Hodge-Laplacian, with we defined in the main text, and the normalized Laplacian, which we discuss below. Additionally, we comment on random walks hypergraphs Laplacians. However, since they...

    work page 2006

  2. [5]

    Collab Imdb Reddit Mutag Enzymes Proteins Peptides-func Peptides-struct # graphs 5000 1000 2000 188 600 1113 15,535 15,535 avg. # node per graph 74.49 19.77 425.57 17.93 31.86 37.40 150.94 150.94 # classes 3 2 2 2 6 2 10 - Table 5: Dataset Statistics for Collab, Imdb, Reddit, Mutag, Enzymes, Proteins and Peptides. 25 C.1.2 Hypergraph Datasets We use five ...

    work page 2000

  3. [6]

    Pubmed Cora-CA Cora-CC Citeseer DBLP # hypernodes, V 19717 2708 2708 3312 43413 # hyperedges, E 7963 1072 1579 1079 22535 # features, d 500 1433 1433 3703 1425 # classes, q 3 7 7 6 6 Table 6: Dataset Statistics C.1.3 BREC Dataset for empirical expressivity analysis The BREC dataset is an expressiveness dataset containing 1-WL-indistinguishable graphs in 4...

    work page 2024

  4. [8]

    Inductive representation learning on large graphs

    Will Hamilton, Zhitao Ying, and Jure Leskovec. Inductive representation learning on large graphs. Advances in neural information processing systems , 30, 2017a. 15 William L. Hamilton, Zhitao Ying, and Jure Leskovec. Inductive Representation Learning on Large Graphs. In NIPS, pages 1024–1034, 2017b. Jing Huang and Jie Yang. Unignn: a unified framework for...

    arXiv

  5. [9]

    Semi-supervised classification with graph convolutional net- works

    Thomas N Kipf and Max Welling. Semi-supervised classification with graph convolutional net- works. arXiv preprint arXiv:1609.02907 ,

    arXiv

  6. [10]

    Hypergraph transformer for semi-supervised classification

    Zexi Liu, Bohan Tang, Ziyuan Ye, Xiaowen Dong, Siheng Chen, and Yanfeng Wang. Hypergraph transformer for semi-supervised classification. In ICASSP 2024-2024 IEEE International Con- ference on Acoustics, Speech and Signal Processing (ICASSP) , pages 7515–7519. IEEE,

    work page 2024

  7. [11]

    On the Expressiveness and Generalization of Hypergraph Neural Networks

    Zhezheng Luo, Jiayuan Mao, Joshua B Tenenbaum, and Leslie Pack Kaelbling. On the expres- siveness and generalization of hypergraph neural networks. arXiv preprint arXiv:2303.05490 ,

    work page Pith review arXiv

  8. [13]

    We can actually see they are the same (up to row permutation. 0.0 0.19583 0.0 0.1916 0.0 0.20555556 0.0 0.18 0.0 0.196 0.0 0.18 0.0 0.196 0.0 0.19583 0.0 0.18 0.0 0.1916 0.0 0.196 0.0 0.196 0.0 0.1916 0.0 0.1916 0.0 0.19583 0.0 0.18 0.0 0.18 0.0 0.18 0.0 0.205 0.0 0.19583 Table 13: Pair A (left) and Pair B (right) 2-R WPE encodings. They match if we reord...

    work page 1916

Show all 29 references
  1. [14]

    Petar Veliˇ ckovi´ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio

    URL https://arxiv.org/abs/2406.06642. Petar Veliˇ ckovi´ c, Guillem Cucurull, Arantxa Casanova, Adriana Romero, Pietro Lio, and Yoshua Bengio. Graph attention networks. arXiv preprint arXiv:1710.10903 ,

    arXiv

  2. [17]

    From stars to subgraphs: Uplifting any gnn with local structure awareness

    17 Lingxiao Zhao, Wei Jin, Leman Akoglu, and Neil Shah. From stars to subgraphs: Uplifting any gnn with local structure awareness. arXiv preprint arXiv:2110.03753 ,

    arXiv

  3. [18]

    36 E.4 Additional Plots Figure 10: The pair 0 of the regular category in BREC

    0.25 0.40824829 0.40824829 0.25 0.40824829 0.40824829 0.25 0.25 0.20412415 0.20412415 0.2773501 0.39223227 0.2773501 0.39223227 0.39223227 0.39223227 0.2773501 0.2773501 0.19611614 0.19611614 Table 18: Pair A (left) and Pair B (right) Normalized H-1-LAPE encodings. 36 E.4 Addi...

    2021

  4. [19]

    and HyperNetX (Praggastis et al., 2023). A.2 Lifting graphs to hypergraphs The term “lifting” refers generally to the reparametrization of one topological domain to another, usually one that captures richer higher-order information. In our setting we lift graphs to hypergraphs...

    2023

  5. [22]

    The probability of a random walk transitioning from node i to j is given by −Lij = Aij di

    20 A.4.2 Random walks hypergraphs Laplacians For a graph, the random walk Laplacian is defined as L = I − D−1A, where, as usual, D denotes the degree matrix and A the adjency matrix. The probability of a random walk transitioning from node i to j is given by −Lij = Aij di . Mu...

    2022

  6. [23]

    B.2.2 UniGIN UniGIN also follows the two-phase scheme (see Eq

    is the average degree of an hyperedge (after adding self-loops to the original hypergraph), and where ˜Ni and ˜E(i) are the neighborhood of vertex i and the incident hyperedges to i after adding self loops. B.2.2 UniGIN UniGIN also follows the two-phase scheme (see Eq. 2.1) an...

    2021

  7. [26]

    We use the same pre-processed hypergraphs as in Yadati et al

    and DBLP (Rossi and Ahmed, 2015). We use the same pre-processed hypergraphs as in Yadati et al. (2019), which are taken from Huang and Yang (2021). The hypergraphs are created with each vertex representing a document. The Cora data set, for example, contains machine learning p...

    2019

  8. [1968]

    How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 ,

    Keyulu Xu, Weihua Hu, Jure Leskovec, and Stefanie Jegelka. How powerful are graph neural networks? arXiv preprint arXiv:1810.00826 ,

    arXiv

  9. [1973]

    typhimurium TA98

    using S. typhimurium TA98. Proteins and Enzymes are introduced in (Borgwardt et al., 2005). These datasets use the 3D structure of the folded proteins to build a graph of amino acids (Borgwardt et al., 2005). Peptides is a chemical data set introduced in (Dwivedi et al., 2022)...

    2005

  10. [1991]

    Graph neural networks with learnable structural and positional representations.arXiv preprint arXiv:2110.07875,

    Vijay Prakash Dwivedi, Anh Tuan Luu, Thomas Laurent, Yoshua Bengio, and Xavier Bresson. Graph neural networks with learnable structural and positional representations.arXiv preprint arXiv:2110.07875,

    arXiv

  11. [2006]

    Here, we focus on clique expansion, which we empirically found to be the best performing expansion

    18 A Extended Background A.1 Hypergraph Expansions There exist several expansion techniques for reparametrizing hypergraphs as graphs. Here, we focus on clique expansion, which we empirically found to be the best performing expansion. For more details see, e.g., (Sun et al., 2...

    2008

  12. [2008]

    Higher-order attention networks

    Mustafa Hajij, Ghada Zamzmi, Theodore Papamarkou, Nina Miolane, Aldo Guzm´ an-S´ aenz, and Karthikeyan Natesan Ramamurthy. Higher-order attention networks. arXiv preprint arXiv:2206.00606, 2(3):4,

    arXiv

  13. [2011]

    A new model for learning in graph do- mains

    Marco Gori, Gabriele Monfardini, and Franco Scarselli. A new model for learning in graph do- mains. In Proceedings. 2005 IEEE international joint conference on neural networks, volume 2, pages 729–734,

    2005

  14. [2015]

    Forman’s ricci curvature-from networks to hypernetworks

    Emil Saucan and Melanie Weber. Forman’s ricci curvature-from networks to hypernetworks. In Complex Networks and Their Applications VII: Volume 1 Proceedings The 7th International Conference on Complex Networks and Their Applications COMPLEX NETWORKS 2018 7 , pages 706–717. Springer,

    2018

  15. [2017]

    Towards better evaluation of gnn expressiveness with brec dataset

    Yanbo Wang and Muhan Zhang. Towards better evaluation of gnn expressiveness with brec dataset. arXiv preprint arXiv:2304.07702 ,

    arXiv

  16. [2019]

    The Dirichlet energy E(f ) of a scalar function on a hypergraph is defined as E(f ) = 1 2 X e∈E X {u,v}⊆e 1 |e| f (u)p d(u) − f (v)p d(v) !2

    is defined as ∆ = I − D−1/2 v B1D−1 e BT 1 D−1/2 v = D−1/2 v (Dv − B1D−1 e BT 1 )D−1/2 v , (A.4) where Dv and De are the diagonal node and edge degree matrices. The Dirichlet energy E(f ) of a scalar function on a hypergraph is defined as E(f ) = 1 2 X e∈E X {u,v}⊆e 1 |e| f (u...

    2006

  17. [2020]

    Ollivier-ricci curvature for hyper- graphs: A unified framework

    Corinna Coupette, Sebastian Dalleiger, and Bastian Rieck. Ollivier-ricci curvature for hyper- graphs: A unified framework. arXiv preprint arXiv:2210.12048 ,

    arXiv

  18. [2021]

    Higher-order network analysis takes off, fueled by classical ideas and new data

    Austin R Benson, David F Gleich, and Desmond J Higham. Higher-order network analysis takes off, fueled by classical ideas and new data. arXiv preprint arXiv:2103.05031 ,

    arXiv

  19. [2022]

    A simple yet effective baseline for non-attributed graph classification

    Chen Cai and Yusu Wang. A simple yet effective baseline for non-attributed graph classification. arXiv preprint arXiv:1811.03508 ,

    arXiv

  20. [2023]

    Hypernetx: A python package for modeling complex network data as hypergraphs

    Brenda Praggastis, Sinan Aksoy, Dustin Arendt, Mark Bonicillo, Cliff Joslyn, Emilie Purvine, Madelyn Shapiro, and Ji Young Yun. Hypernetx: A python package for modeling complex network data as hypergraphs. arXiv preprint arXiv:2310.11626 ,

    arXiv

  21. [2024]

    R. Forman. Bochner’s method for cell complexes and combinatorial Ricci curvature. Discrete and Computational Geometry , 29(3):323–374, 2003a. Robin Forman. Bochner’s method for cell complexes and combinatorial ricci curvature. Discrete & Computational Geometry , 29:323–374, 20...

    arXiv

Pith tools

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