acceptodds
Under review as a conference paper at ICLR 2027

Generative Modeling of Hypergraphs using Latent Hyperbolic Geometry

Abstract

Hypergraphs encode multi-way interactions among entities arising in a broad range of real-world relational systems. Generative modeling of hypergraphs aims to capture patterns in these interactions for understanding and simulating complex relational structures. Existing approaches often face a difficult trade-off: parsimonious probabilistic models may oversimplify rich multi-way interactions, whereas expressive deep neural architectures can sacrifice interpretability and scalability. To strike a better balance between parsimony and expressivity, we introduce a novel generative model, the *Hyperbolic Hypergraph Generator*, which harnesses hyperbolic geometry as a source of expressive power while retaining a parsimonious and interpretable probabilistic latent space formulation. Theoretically, we establish identifiability of the latent geometry under a geometric compatibility condition, and characterize the gain in expressivity afforded by the hyperbolic latent space. For training and sampling, we develop *H2G*, a scalable procedure that combines likelihood-based embedding with latent space generator construction, and establish theoretical guarantees for the consistency of the generated hyperlink distribution. Extensive empirical comparisons with baseline methods show that H2G captures meaningful latent geometry and delivers strong generative performance, achieving leading results on fidelity metrics that measure the reproduction of key structural characteristics of real-world hypergraphs. Together, these theoretical and empirical results demonstrate that our model achieves an effective balance between parsimony and expressivity, while highlighting the broader potential of hyperbolic geometry for modeling complex multi-way relational systems.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hypergraphs, Hyperbolic Geometry, Generative Models, Latent Space Methods

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_2TRZOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes H2G (Hyperbolic Hypergraph Generator), a hypergraph generation framework based on latent hyperbolic geometry. The method first learns hyperbolic embeddings of nodes and hyperedges, node popularity parameters, and spatial curvature through maximum likelihood estimation. It then generates new latent hyperedge embeddings using either empirical resampling (H2G-EB) or a Riemannian diffusion model (H2G-RD). Finally, new hyperedges are generated through probabilistic decoding based on hyperbolic distances and node popularity. The paper also analyzes curvature identifiability, higher-order dependencies, and consistency of the generated distribution, and demonstrates the effectiveness of the proposed method on three real-world hypergraph datasets.

Critical strengths
  1. Interpretable Generative Framework with Theoretical Support. The paper combines hyperbolic geometry with probabilistic generative modeling, capturing complex multi-node interactions through low-dimensional embeddings, node popularity, and geometric distances. The framework is relatively simple and interpretable. Furthermore, the theoretical analyses of curvature identifiability, higher-order dependencies, and generative consistency provide solid theoretical support for the proposed method.

  2. Strong Empirical Performance. The experiments cover three real-world hypergraph datasets and include comparisons with various generative models. H2G achieves competitive results on generation quality metrics such as Energy Distance and MMD², while using only a two-dimensional hyperbolic latent space, demonstrating strong representational capability and practical potential.

Critical weaknesses
  1. Limited Scalability Evaluation. The paper emphasizes the scalability of the proposed method, but the experiments mainly focus on small- to medium-scale hypergraphs. Although some computational cost information is provided, a systematic efficiency evaluation across different dataset scales and runtime comparisons with baselines are still lacking. Given that the method involves curvature grid search, manifold optimization, and diffusion training, its computational efficiency on larger hypergraphs remains to be validated.

  2. Insufficient Validation of Hyperbolic Geometry. The paper does not sufficiently establish the independent contribution of hyperbolic geometry. The existing DDE baseline uses Euclidean inner products, whereas H2G employs hyperbolic distances, changing both the geometric space and the interaction function. This makes it difficult to isolate the actual benefits of negative curvature. Furthermore, a systematic analysis of the conditions under which hyperbolic geometry is advantageous is lacking.

Questions for authors
  1. Algorithm 1 requires separate embedding optimization for multiple candidate curvature values, while Appendix B.7 only reports the computational cost on MIMIC-III under a single fixed curvature. Could the authors clarify the total training cost of the full curvature search and how the computational overhead scales with the number of nodes and hyperedges?

  2. Appendix D.3 compares DDE with different Euclidean embedding dimensions, but its inner-product-based interaction mechanism differs from H2G's distance-based decoder. Have the authors considered a Euclidean model with the same Bernoulli distance-based decoder? Could such a comparison help isolate the contribution of negative curvature to generation quality?

4 comments

Thank you for taking the time and effort to review our paper and your constructive suggestions. We appreciate your recognition of our theoretical and empirical contributions. Below, we address your concerns and questions point by point.

W1 & Q1. Regarding computational efficiency, full curvature search, and scalability

Response:

Implementation improvements. Since submission, we have improved computational efficiency by using a single inner MGD epoch, fewer embedding-fitting iterations, and a coarser curvature grid, while maintaining comparable generation fidelity (main results are updated in Section 4.2). We also chunk the probability matrices and use compressed sparse row (CSR) storage for the incidence data to reduce GPU memory consumption, particularly for large sparse hypergraphs.

We summarize the results of our newly added experiments below, with further details in Appendix B.7 of our revised manuscript.

1. Large-scale feasibility.

We further evaluate H2G on two large-scale hypergraphs, using a lightweight initialization to reduce computational overhead while keeping the model, likelihood objective, and optimization procedure unchanged.

Real-world hypergraph. We trained H2G on threads-ubuntu [1, 2], a real-world hypergraph with 90,054 nodes, 115,987 hyperlinks, and an average hyperlink size of 2.30. We then used the fitted model to generate synthetic hyperlinks on the same node set.

Synthetic hypergraph. We trained H2G on a bipartite stochastic block model [3] with 32,768 nodes, 32,768 hyperlinks, a 3×33\times3 block interaction structure, and an expected hyperlink size of 5. H2G generated 1,048,576 hyperlinks, 32 times the number in the training hypergraph. The generated hypergraph preserved the block structure, with estimated block-wise connection probabilities close to the ground truth.

We refer the reviewer to Appendix B.7 of the revised manuscript for the full experimental setup and computational details. These experiments demonstrate the feasibility of H2G on substantially larger hypergraphs than those considered in the original manuscript.

2. Scalability with hypergraph size. At a fixed latent dimension, each full embedding update costs O(nm)O(nm), and the optimization cost scales with the number of curvature candidates and updates per candidate. With a fixed batch size and architecture, each diffusion training step has no direct dependence on nn or mm. Decoding m~\tilde m generated hyperlinks costs O(m~n)O(\tilde m n).

Appendix B.7.3 reports runtime and peak GPU memory as we vary nn and mm. H2G-RD scales across the tested configurations, with embedding fitting accounting for most of the runtime. Table 9 separates the memory cost of initialization from that of embedding optimization.

3. Cost of full curvature search and comparison with baselines. Appendix B.7 of the revised manuscript reports end-to-end runtime and peak GPU memory for H2G's full curvature search and the baselines on all three real-world datasets. The comparison uses three curvature candidates for H2G and three hyperparameter configurations for each baseline.

References:

[1] Benson, A. R., Abebe, R., Schaub, M. T., Jadbabaie, A., and Kleinberg, J. (2018). Simplicial closure and higher-order link prediction. Proceedings of the National Academy of Sciences, 115(48), E11221–E11230.

[2] Lee, G., Choe, M., and Shin, K. (2021). How do hyperedges overlap in real-world hypergraphs? Patterns, measures, and generators. Proceedings of the Web Conference 2021, 3396–3407.

[3] Coupette, C., Dalleiger, S., and Rieck, B. (2023). Ollivier-Ricci curvature for hypergraphs: A unified framework. International Conference on Learning Representations (ICLR).

W2 & Q2. Regarding the controlled Euclidean comparison

Response:

Thank you for this suggestion. We now have introduced the Euclidean Hypergraph Generator (EHG), which uses the same Bernoulli distance-based decoder as H2G with Euclidean distances:

\qquad f,z_i\in\mathbb{R}^d.$$ DDE uses Euclidean inner products, so comparing it with H2G changes both the latent geometry and the interaction function. EHG lets us compare the two geometries under the same decoder. **1. Experiments.** We implement EHG-EB and EHG-D as the Euclidean counterparts of H2G-EB and H2G-RD, respectively. The implementations differ in geometry-dependent initialization and optimization (USVT+EU20 and MGD for H2G, USVT and GD for EHG) and in the diffusion process for the diffusion variants. We focus on H2G-EB and EHG-EB so that the comparison uses the same empirical bootstrap procedure in both geometries. We select the Euclidean dimension from $d\in\\{2,4,8\\}$ using validation Energy, the same criterion used to select curvature for H2G. **Real data.** Appendix D.9 reports the full comparisons. The table below shows test means (SDs) for the selected configurations. Energy is in units of $10^{-3}$ for MIMIC-III and $10^{-4}$ for MathOverflow-Answers and CORUM. MMD\$^2\$ is in units of $10^{-3}$ throughout. | Dataset | EHG dimension | H2G-EB Energy | EHG-EB Energy | H2G-EB MMD² | EHG-EB MMD² | |---|---:|---:|---:|---:|---:| | MIMIC-III | 8 | 1.34 (0.07) | **1.17 (0.105)** | 7.45 (0.05) | **7.34 (0.114)** | | MathOverflow-Answers | 8 | **3.67 (0.36)** | 9.92 (1.1) | **3.52 (0.02)** | 3.93 (0.18) | | CORUM | 2 | **7.33 (0.60)** | 41.3 (2.4) | **2.16 (0.09)** | 4.41 (0.30) | EHG-EB performs slightly better on MIMIC-III, while H2G-EB achieves substantially lower Energy and MMD\$^2\$ on MathOverflow-Answers and CORUM. **Synthetic data.** Appendix C.4 compares both geometries in a controlled $2\times2$ study varying the ground-truth geometry and average hyperlink order. Both fitted models use two-dimensional embeddings. Test results are means (SDs) in units of $10^{-3}$. | True geometry | Mean order | H2G-EB Energy | EHG-EB Energy | H2G-EB MMD² | EHG-EB MMD² | |---|---:|---:|---:|---:|---:| | Hyperbolic | 5.03 | **3.716 (0.932)** | 7.906 (1.721) | **9.129 (0.651)** | 11.574 (0.611) | | Hyperbolic | 19.96 | **2.438 (0.967)** | 2.439 (0.996) | **7.876 (0.383)** | 7.878 (0.334) | | Euclidean | 4.99 | **0.605 (0.546)** | 1.192 (0.843) | **7.165 (0.419)** | 7.415 (0.455) | | Euclidean | 19.97 | **0.699 (0.464)** | 0.846 (0.493) | **7.146 (0.152)** | 7.202 (0.141) | Under sparse hyperbolic ground truth, H2G-EB reduces Energy by **53.0%** and MMD\$^2\$ by **21.1%**, while the two models perform nearly identically at the higher hyperlink order. H2G-EB also achieves lower mean Energy and MMD\$^2\$ at both orders under Euclidean ground truth. These findings suggest that hyperbolic geometry can offer advantages in hyperlink generation without requiring the true latent geometry to be strictly hyperbolic. **2. Theory.** Hyperbolic geometry does not consistently outperform Euclidean geometry, and its advantages depend on the underlying hyperlink distribution. Beyond the tree-metric representation result in Proposition A.1, we have added Propositions 2–3 (Section 2.3, with proofs in Appendix E.4) to directly compare the expressivity of the induced hyperlink distributions. Proposition 2 establishes that a two-dimensional hyperbolic model can approximate certain overlapping modular hyperlink distributions arbitrarily well, whereas Euclidean models of any fixed maximum dimension have a nonvanishing total variation (TV) approximation gap. Conversely, Proposition 3 constructs a distribution over three pairwise hyperlinks that Euclidean models can approximate arbitrarily well, while hyperbolic models with curvature bounded away from zero have a strictly positive TV approximation gap. These results show that neither geometry is universally superior. In practice, hypergraph structures are typically more complex, making it difficult to choose a latent geometry solely by looking for patterns like those in the propositions. A held-out validation set can help guide this choice by comparing the generative fidelity of the fitted models. In summary, we introduce H2G, provide an accompanying toolkit, and examine how its expressivity and performance compare with Euclidean alternatives. In practice, practitioners should choose the model that best fits their application, and our results can help guide that choice. Thank you again for your helpful feedback. Please let us know if you have any additional questions. Meanwhile, we are continuing to proofread the manuscript and double-check the details.

Thanks for the author's rebuttal response. I maintain my score.

Thank you very much for taking the time to read our rebuttal response. We appreciate your careful review and feedback.

Reviewer_NNi7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a hypergraph generation framework H2G based on latent hyperbolic geometry, which recovers the potential positions of nodes and hyperedges through maximum likelihood estimation, and then uses empirical resampling or Riemann diffusion to learn the hyperedge embedding distribution, ultimately generating new hyperedges; Its theoretical contributions focus on curvature recognizability and generative consistency.

Critical strengths
  1. This paper introduces H2G, a unified framework for modeling hyperedge generation through latent hyperbolic geometry. By representing nodes and hyperedges in hyperbolic space, this method provides a meaningful approach to capture structural dependencies.

  2. This paper theoretically investigates the model identifiability, including curvature and latent geometry, characterizes the induced higher-order dependencies of the model, and compares the expressivity of two-dimensional hyperbolic space with that of Euclidean spaces of increasing dimension.

Critical weaknesses
  1. This paper indicates that hyperlinks generated by H2G are rarely exact copies of the training set, and H2G-RD can sample potential embeddings beyond training dataset support. However, these only confirm that the generated hyperlinks are different from the observed training hyperlinks; It has not been directly determined whether the new combination comes from a reasonable sample of the underlying real-world hyperlink distribution. A stronger evaluation of invisible hyperlinks, such as targeting real hyperlinks or validating them through specific domains, will better support the claim of generating meaningful, previously unseen hyperlinks.

  2. This paper partially attributes the performance advantage of H2G to the expressive ability of hyperbolic latent space, but there seems to be a lack of a strictly controlled comparison in the experiment: only changing the geometric structure of the latent space (such as changing from Euclidean space to hyperbolic space) while maintaining consistency in other factors such as embedding form, optimization process, diffusion dynamics, and sampling flow. Therefore, the current experimental results cannot clearly prove whether the observed performance gain mainly comes from the core proposition of the paper, which is to introduce hyperbolic geometry itself.

Questions for authors
  1. Why does H2G-RD underperform H2G-EB across all reported metrics on MathOverflow-Answers?

  2. How sensitive is the model to curvature estimation and the choice of the curvature grid?

  3. Does hyperbolic geometry consistently outperform comparable Euclidean latent-space models?

  4. How does the computational cost of H2G-RD scale with the number of nodes and hyperedges?

4 comments

Thank you for taking the time and effort to review our paper and for your constructive questions. Below, we address your concerns and questions point by point.

W1. Regarding the validity of generated hyperlinks

Response:

We agree that novelty alone does not establish that generated hyperlinks are realistic. A generated hyperlink can match one seen in training, but such matches may be rare when probability mass is spread across many of the 2n2^n possible node combinations. We therefore assess whether generated hyperlinks exhibit plausible domain-specific relationships and reproduce key distributional characteristics of real hyperlinks.

Domain-specific validation. Following your suggestion, we have added a qualitative analysis of MathOverflow-Answers in the revised manuscript (Appendix D.6.3). Both hyperlinks connect questions spanning closely related mathematical fields, while exhibiting distinct topic combinations. This provides qualitative evidence that H2G can generate plausible, previously unseen user–question interaction patterns.

The original manuscript also provides domain-specific interpretations: on MIMIC-III, Figure 1 illustrates a generated patient record with heart conditions and clinically related symptoms (Introduction), while Figure 3 reveals interpretable latent symptom clusters (Section 4.2 and Appendix D.5.3); on CORUM, Figure 4 presents generated protein complexes associated with gene regulation, cell adhesion, and DNA repair (Appendix D.7.3).

Distributional fidelity. Given the sparsity and combinatorial nature of hypergraphs, exact matches between generated and held-out hyperlinks are rare. Our goal is therefore to recover the underlying hyperlink distribution rather than individual hyperlinks. In Section 4.2 of our manuscript, we use Energy distance and MMD$^2$ to compare the overall distributions of real and generated hyperlinks, based on Jaccard distance and a Jaccard-based RBF kernel, respectively. We also evaluate node frequencies and co-occurrence statistics using ΔHd\Delta_{H_d} and ΔHv\Delta_{H_v}. The results show that H2G closely reproduces these distributional characteristics.

Together, these results provide statistical and domain-specific evidence for meaningful hyperlink generation.

W2. Regarding the independent contribution and advantages of hyperbolic geometry

Response:

Thank you for this valuable suggestion. We agree that the comparison with DDE does not fully isolate the contribution of hyperbolic geometry, since their interaction functions differ. To address this, we introduce the Euclidean Hypergraph Generator (EHG), which directly replaces hyperbolic distances with Euclidean distances while retaining the same Bernoulli distance-based decoder:

$ p_i(f)=\sigma\left(\alpha_i-\lVert f-z_i\rVert_2\right), \qquad f,z_i\in\mathbb{R}^d. $

We implement EHG-EB as the Euclidean counterpart of H2G-EB, respectively. Geometry-dependent differences remain in initialization (including USVT+EU20 versus USVT), embedding optimization (MGD versus GD), and diffusion (Riemannian versus Euclidean), while all other training settings are kept identical. We consider d∈{2,4,8}d\in\{2,4,8\} and select the dimension using validation Energy, the same criterion of curvature selection for H2G.

The tables below report part of the results, with full dimension comparisons included in Appendix D.9 of the revised manuscript.

MIMIC-III. EHG-EB achieves slightly lower test Energy and MMD$^2$ than H2G-EB at d=8d=8, suggesting that Euclidean geometry can also provide an effective representation for this dataset.

MethodEnergy (×10−3\times 10^{-3})MMD$^2$ (×10−3\times 10^{-3})
H2G-EB1.34 (0.07)7.45 (0.05)
EHG-EB (d=8d=8)1.17 (0.105)7.34 (0.114)

MathOverflow-Answers. H2G-EB substantially outperforms EHG-EB in test Energy and MMD$^2$, even with eight-dimensional Euclidean embeddings.

MethodEnergy (×10−4\times 10^{-4})MMD$^2$ (×10−3\times 10^{-3})
H2G-EB3.67 (0.36)3.52 (0.02)
EHG-EB (d=8d=8)9.92 (1.1)3.93 (0.18)

CORUM. H2G-EB achieves substantially lower test Energy and MMD$^2$ than EHG-EB at its selected dimension d=2d=2, demonstrating the advantage of hyperbolic geometry in capturing the hyperlink distribution.

MethodEnergy (×10−4\times 10^{-4})MMD$^2$ (×10−3\times 10^{-3})
H2G-EB7.33 (0.60)2.16 (0.09)
EHG-EB (d=2d=2)41.3 (2.4)4.41 (0.30)

Overall, these controlled comparisons show that hyperbolic geometry does not universally outperform Euclidean geometry, but provides substantial improvements in distributional fidelity on MathOverflow-Answers and CORUM.

Meanwhile, as discussed in our response to Reviewer nQV4, our new Propositions 2–3 (in the new Section 2.3) give examples favoring each geometry. H2G adds a hyperbolic option to the toolkit, and practitioners can choose the geometry best suited to their data.

Q1. Why does H2G-RD underperform H2G-EB across all reported metrics on MathOverflow-Answers?

Response:

Thank you for this observation. We initially limited H2G-RD's compute budget, leaving little room to tune its diffusion stage; while H2G-EB's bootstrap step requires no tuning. We did not tune H2G-RD extensively at first because H2G-EB already outperformed most baselines. With faster embedding fitting (see Q4) and more compute for tuning, H2G-RD now performs more comparably to H2G-EB.

Updated generation fidelity on MathOverflow-Answers (mean (SD) over five sampling seeds, lower is better).

MethodEnergy (×10−4\times 10^{-4})MMD$^2$ (×10−3\times 10^{-3})ΔHd\Delta_{H_d} (×10−3\times 10^{-3})ΔHv\Delta_{H_v} (×10−4\times 10^{-4})
H2G-EB3.67 (0.36)3.52 (0.02)4.09 (0.01)4.08 (0.00)
H2G-RD4.27 (0.31)3.51 (0.02)4.10 (0.01)4.09 (0.00)

H2G-RD’s weaker initial performance may reflect its additional diffusion step on the fitted hyperlink embeddings. This step offers greater flexibility in learning the latent distribution but can introduce approximation error and is sensitive to hyperparameter tuning.

Q2. How sensitive is the model to curvature estimation and the choice of the curvature grid?

Response:

We extend the simulation analysis of curvature sensitivity and grid selection to all three real-world datasets.

Synthetic experiments. As shown in Appendix C.2.5, when the true curvature is K∗=0.5K^*=0.5, the training log-likelihood generally decreases and validation Energy (the Energy distance between generated and real validation hyperlinks, lower is better) increases as the candidate curvature moves away from the true value. Both criteria attain their optima at K=0.5K=0.5. These results are consistent with our theoretical motivation for curvature identifiability and demonstrate the effect of curvature misspecification on generation quality.

Curvature grid selection. For all three real-world datasets (MIMIC-III, MathOverflow-Answers, and CORUM), we select curvature by validation Energy using a coarse-to-fine search. We start from K={0.1,0.3,1.0}\mathcal{K}=\{0.1,0.3,1.0\}, then optionally refine the grid around an interior optimum or expand it when the optimum lies at a boundary, followed by further refinement as needed.

As discussed in Q1, H2G-RD introduces additional diffusion approximation error and is more sensitive to hyperparameter tuning. We therefore focus on H2G-EB to better isolate the effect of curvature on generation quality. All results below are reported as mean (SD) of validation Energy (lower is better). Further details are provided in Appendix D.8 of the revised manuscript.

MIMIC-III. The original grid selects K=0.3K=0.3. After local refinement, K=0.2K=0.2 achieves a slightly lower validation Energy (0.971 versus 0.972). Performance remains relatively stable near the selected curvature, suggesting limited sensitivity within this favorable range.

KK0.10.20.30.51.0
Energy (×10−3\times 10^{-3})1.04 (0.11)0.971 (0.11)0.972 (0.08)1.13 (0.05)1.07 (0.09)

MathOverflow-Answers. After refinement, K=0.3K=0.3 remains optimal for H2G-EB within the tested grid. Performance deteriorates as curvature moves away from this value, particularly at K=0.1K=0.1, suggesting greater curvature sensitivity than on MIMIC-III.

KK0.10.20.30.51.0
Energy (×10−4\times 10^{-4})16.3 (1.9)3.87 (0.31)3.08 (0.19)3.47 (0.29)4.70 (0.17)

CORUM. The original coarse grid selects K=1.0K=1.0, which already achieves good generation fidelity. Expanding the grid further reduces validation Energy, reaching its minimum at K=3.0K=3.0 before increasing again at K=5.0K=5.0 and 10.010.0. This suggests a favorable curvature range around K=3.0K=3.0 and demonstrates the benefit of grid refinement.

KK0.10.30.51.01.53.05.010.0
Energy (×10−4\times 10^{-4})12.4 (0.84)10.5 (1.5)9.59 (0.96)7.03 (0.68)6.53 (1.1)5.68 (0.81)6.15 (0.68)7.66 (1.4)

A practical strategy is to start with a broad, coarse grid, expand it if the best candidate lies at a boundary, and refine around promising values. In our experiments, one or two rounds of expansion or refinement were enough to find a good KK, with nearby values giving similar validation Energy. Further refinement is only worthwhile if the expected gain justifies the added compute. We use the expanded grid for sensitivity analysis (Appendix D.8) and retain the original search budget for the main benchmarks.

Q3. Does hyperbolic geometry consistently outperform comparable Euclidean latent-space models?

Response:

Thank you for this insightful question. Hyperbolic geometry does not consistently outperform Euclidean geometry, and its advantages depend on the underlying hyperlink distribution. Beyond the tree-metric representation result in Proposition A.1, we have added Propositions 2–3 (Section 2.3, with proofs in Appendix E.4) to directly compare the expressivity of the induced hyperlink distributions. Proposition 2 establishes that a two-dimensional hyperbolic model can approximate certain overlapping modular hyperlink distributions arbitrarily well, whereas Euclidean models of any fixed maximum dimension have a nonvanishing total variation (TV) approximation gap. Conversely, Proposition 3 constructs a distribution over three pairwise hyperlinks that Euclidean models can approximate arbitrarily well, while hyperbolic models with curvature bounded away from zero have a strictly positive TV approximation gap. These examples show that the preferred geometry depends on the co-occurrence pattern in the hypergraphs.

In practice, hypergraph structures are typically more complex, making it difficult to choose a latent geometry solely by looking for patterns like those in the propositions. A held-out validation set can help guide this choice by comparing the generative fidelity of the fitted models.

We further conduct a controlled 2×22\times2 synthetic study varying the ground-truth geometry (hyperbolic versus Euclidean) and average hyperlink order. Details are provided in Appendix C.4 of our revised manuscript. H2G-EB and EHG-EB both use two-dimensional latent embeddings. The following tables report test-set means (SDs).

Hyperbolic ground truth — Relatively sparse (average hyperlink order = 5.03)

MethodEnergy (×10−3\times 10^{-3})MMD$^2$ (×10−3\times 10^{-3})
H2G-EB3.716 (0.932)9.129 (0.651)
EHG-EB7.906 (1.721)11.574 (0.611)

Hyperbolic ground truth — Relatively dense (average hyperlink order = 19.96)

MethodEnergy (×10−3\times 10^{-3})MMD$^2$ (×10−3\times 10^{-3})
H2G-EB2.438 (0.967)7.876 (0.383)
EHG-EB2.439 (0.996)7.878 (0.334)

Euclidean ground truth — Relatively sparse (average hyperlink order = 4.99)

MethodEnergy (×10−3\times 10^{-3})MMD$^2$ (×10−3\times 10^{-3})
H2G-EB0.605 (0.546)7.165 (0.419)
EHG-EB1.192 (0.843)7.415 (0.455)

Euclidean ground truth — Relatively dense (average hyperlink order = 19.97)

MethodEnergy (×10−3\times 10^{-3})MMD$^2$ (×10−3\times 10^{-3})
H2G-EB0.699 (0.464)7.146 (0.152)
EHG-EB0.846 (0.493)7.202 (0.141)

Under hyperbolic ground truth with relatively sparse interactions, H2G-EB reduces Energy by 53.0% and MMD$^2$ by 21.1% compared with EHG-EB, while the two methods achieve nearly identical performance at the higher hyperlink order. Notably, even under Euclidean ground truth, where the hyperbolic geometry is misspecified, H2G-EB remains competitive, achieving lower Energy and MMD$^2$ at both hyperlink orders. These findings suggest that hyperbolic geometry can offer advantages in hyperlink generation without requiring the true latent geometry to be strictly hyperbolic.

Q4. How does the computational cost of H2G-RD scale with the number of nodes and hyperedges?

Response:

Thank you for your valuable question regarding computational scalability. We provide additional theoretical complexity analysis and empirical results on runtime and GPU memory scaling, demonstrating the computational feasibility of H2G on larger hypergraphs. Details are provided in Appendix B.7 of our revised manuscript, where we further compare the computational cost of H2G with baselines and run experiments on substantially larger hypergraphs, generating over 1 million hyperlinks while retaining the data structure.

Theoretical complexity. For a fixed latent dimension and training budget, embedding fitting requires O(nm)O(nm) operations per full update, as the likelihood involves all hyperlink–node pairs. The total fitting cost scales linearly with the number of curvature candidates and optimization updates. Diffusion training operates on latent hyperlink embeddings and has no direct dependence on nn or mm per training step. Generating m~\tilde m hyperlinks requires O(m~n)O(\tilde m n) operations for Bernoulli decoding, in addition to diffusion sampling, whose per-hyperlink cost is independent of nn and mm.

Scalability. Table 9 in Appendix B.7.3 reports runtime and peak GPU memory as we vary the numbers of nodes and hyperlinks. H2G-RD scales across the tested sizes, with embedding fitting accounting for most of the runtime and diffusion and generation adding relatively little overhead.

Thank you again for your feedback. Please let us know if you have any further questions. Meanwhile, we are continuing to proofread the manuscript and double-check the details.

Reviewer_nQV4OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the trade-off between expressivity and interpretability in hypergraph generation. It proposes a Hyperbolic Hypergraph Generator that represents nodes and hyperlinks in a two-dimensional hyperbolic space with learnable curvature. Given a latent hyperlink position, nodes are included independently according to probabilities determined by geometric distance and node-specific popularity. Marginalizing over the shared position induces dependencies among node inclusions. The H2G procedure first estimates embeddings, popularity parameters, and curvature through likelihood-based fitting, then learns the hyperlink-embedding distribution using empirical bootstrap or Riemannian diffusion. Samples from this distribution are decoded into new hyperlinks on the same node set. The paper also provides conditional identifiability and generative-consistency analyses and a geometric motivation for the latent space. Synthetic and real-data studies report strong generation fidelity and interpretable latent organization.

Critical strengths
  1. The paper addresses a meaningful modeling problem: representing heterogeneous multi-way interactions while retaining a transparent generation mechanism. Hyperbolic geometry provides a plausible inductive bias for branching and hierarchical organization, and the two-dimensional construction keeps the geometric representation compact. Separating node popularity from distance makes their respective contributions to inclusion probabilities explicit. The shared hyperlink latent variable induces dependence through marginalization, while the Bernoulli decoder directly accommodates variable hyperlink sizes and repeated interactions. These choices form a clear and coherent probabilistic model.

  2. H2G separates likelihood-based embedding from latent-distribution learning, allowing the same decoder to work with either empirical bootstrap or Riemannian diffusion. This modular design offers both a simple sampling strategy and a learned geometric sampler. The theoretical analysis explains how parameter estimation and latent-distribution recovery affect the generated hyperlink law. In particular, the Wasserstein coupling argument accommodates empirical bootstrap without requiring a finite KL divergence between continuous and discrete latent laws. The explicit distinction between normalized KL consistency and stronger distributional guarantees also makes the scope of the analysis clear.

Critical weaknesses
  1. Theorem 1 fixes node effects and assumes a holomorphic compatibility relation between candidate embedding configurations. The practical procedure jointly fits node effects and embeddings without enforcing that relation. Appendix E.4.1 also introduces an additional quantitative identifiability condition. These qualifications are stated, but their implications for interpreting the fitted curvature could be explained more directly. In particular, the paper should distinguish curvature as a useful model-selection parameter from curvature as a uniquely recoverable property of the data-generating process.

  2. Proposition A.1 establishes an advantage for representing branching-tree metrics, whereas the generator uses hyperlink-node distances, popularity parameters, and a latent mixing distribution. The metric result therefore motivates the geometry but does not fully characterize the family of hyperlink distributions that the model can represent. A clearer account of which interaction patterns suit the two-dimensional construction, and which may require greater dimension or a different geometry, would sharpen the method’s scope and strengthen the motivation.

Questions for authors

see Weaknesses

1 comment

Thank you for taking the time and effort to review our paper and for your insightful comments. We appreciate your recognition of the problem we solve and the positive assessment of the strategies we introduce. In response to your questions and comments on weaknesses, we have substantially strengthened both the empirical and theoretical components of the paper. Below are our point-by-point responses to your comments:

W1. Curvature identifiability and estimation

1. Assumptions and empirical recovery. Prior work on curvature identifiability in hyperbolic latent-space models for relational data has focused on pairwise networks. The closest work, Li et al. [1], assumes holomorphic compatibility for binary pairwise networks without node-specific degree parameters. We extend the roadmap to hypergraphs and introduce node degree parameters for capturing node degree heterogeneity. The current theoretical analysis in Theorem 1 treats these effects as fixed and retains a similar holomorphic compatibility assumption.

Meanwhile, our numerical simulation provides supplement evidence on the identifiability claim. In the simulations, degree parameters are estimated freely without any further restrictions and can be estimated consistently, as shown in Appendix C.2.5, Figure 9 and Table 4, and node-effect and embedding errors decrease with sample size (Appendix C.2.4, Figs. 7–8). This provides empirical evidence of joint recovery, where identifiability is a necessary component, under the simulated designs.

Condition 2 (Appendix E.5.1) directly assumes quantitative identifiability of the full parameter vector after aligning the embeddings by a common isometry. We acknowledge that our curvature-identifiability proof in Theorem 1 requires fixed node effects, as the closest prior work, Li et al. [1], studies a model without node-specific degree effects. Proposition E.2 thus takes a different starting point by assuming joint identifiability under Condition 2 and deriving conditional MLE consistency with freely estimated node effects.

2. Selecting curvature. To maximize the training likelihood over the grid, we fit the other parameters for each candidate KK and select the value of KK with the highest training log-likelihood. We can also treat KK as a hyperparameter and select it using the validation Energy to optimize generative fidelity. The simulation study in Appendix C.2.5 provides empirical evidence that both strategies can recover the true KK. We have revised the manuscript to clarify how the strategies are related and how their objectives differ.

W2. Hyperbolic versus Euclidean

1. Theory. We now add a new Section 2.3 that compares marginal hyperlink laws induced by Euclidean versus hyperbolic latent geometry. Proposition 2 constructs an overlapping modular hypergraph that a 2D hyperbolic model approximates arbitrarily well, while for any fixed-dimensional Euclidean counterpart, a positive TV gap persists in fitting such hypergraph distribution. Proposition 3 gives a reverse example when Euclidean geometry is more suitable than hyperbolic. These examples show that the preferred geometry depends on the co-occurrence pattern in the hypergraphs. In practice, hypergraph structures are typically more complex, making it difficult to choose a latent geometry solely by looking for patterns like those in the propositions. A held-out validation set can help guide this choice by comparing the generative fidelity of the fitted models.

2. Experiments. We also add Appendix C.4 that compares both geometry in a 2×22\times2 study. Test results are mean (SD), in units of 10−310^{-3}. Lower is better.

True geometryMean orderH2G-EB EnergyEHG-EB EnergyH2G-EB MMD²EHG-EB MMD²
Hyperbolic5.033.716 (0.932)7.906 (1.721)9.129 (0.651)11.574 (0.611)
Hyperbolic19.962.438 (0.967)2.439 (0.996)7.876 (0.383)7.878 (0.334)
Euclidean4.990.605 (0.546)1.192 (0.843)7.165 (0.419)7.415 (0.455)
Euclidean19.970.699 (0.464)0.846 (0.493)7.146 (0.152)7.202 (0.141)

H2G-EB achieves lower mean Energy and MMD² at both relatively dense and sparse hypergraphs under Euclidean ground truth, suggesting robustness to misspecification of the latent geometry in these settings. We refer the reviewer to Appendix C.4 for details.

In summary, we introduce H2G, provide an accompanying toolkit, and study how its expressivity and performance compare with Euclidean alternatives. In practice, practitioners should choose the model that best fits their application, and our results could guide that choice.

Thank you again for your feedback. Please let us know if you have any further questions. Meanwhile, we are continuing to proofread the manuscript and double-check the details.

[1] Li, J., Xu, G., and Zhu, J. (2026). Hyperbolic network latent space model with learnable curvature. JASA (2026+).

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to Reviewers

We sincerely thank the Area Chair for their time and consideration, and the reviewers for their constructive feedback and insightful questions. We are encouraged that the reviewers recognized the value of using hyperbolic latent geometry for generative modeling of hypergraphs, together with the theoretical support and strong empirical performance of our work. To address each of the reviewers' questions and concerns, we have provided point-by-point responses and uploaded a revised manuscript with all major changes highlighted in blue.

For conciseness, we refer to Reviewers nQV4, NNi7, and 2TRZ as R1, R2, and R3, respectively, with W and Q denoting weaknesses and questions. Please find the summary of our responses to reviewers' comments below:

  • Expressivity and Comparison of Latent Geometries [R1W2; R2W2; R2Q3; R3W2; R3Q2]

    • We added two propositions that establish complementary expressivity advantages of hyperbolic and Euclidean latent geometries for different hyperlink distributions, demonstrating that the preferred geometry depends on the co-occurrence pattern in the hypergraphs (Section 2.3).
    • We added controlled comparisons between H2G and its Euclidean counterpart EHG using the same Bernoulli distance-based decoder, demonstrating the benefits and limitations of hyperbolic geometry on real-world hypergraphs (Appendix D.9).
    • We also conducted synthetic experiments to compare the performance of both models under hyperbolic and Euclidean ground-truth geometries, suggesting H2G's robustness to misspecification of latent geometry in the reported settings (Appendix C.4).
  • Curvature Identifiability and Sensitivity [R1W1; R2Q2]

    • We clarified the assumptions of Theorem 1 (Section 2.2) and the role of Condition 2 in establishing conditional MLE consistency (Appendix E.5.1), and revised the manuscript to distinguish likelihood-based curvature estimation from validation-based curvature selection (Appendix B.1; Appendix C.2.5).
    • We extended the simulation analysis of curvature sensitivity and grid selection to all three real-world datasets, providing empirical support for our grid selection strategy (Appendix D.8).
  • Computational Efficiency and Scalability [R2Q4; R3W1; R3Q1]

    • We improved embedding-fitting efficiency through fewer optimization iterations, chunked probability matrices, and CSR storage. The fidelity results are updated in Section 4.2, and implementation details in Appendices B.2 and B.4.
    • We analyzed H2G's stage-wise computational complexity, with an O(nm) cost per embedding update, and conducted additional experiments on runtime and GPU memory scaling with hypergraph size (Appendix B.7.2).
    • We additionally reported runtime and GPU memory comparisons with baselines on all three real-world hypergraphs, while accounting for the full curvature search (Appendix B.7.1).
    • We also demonstrated the feasibility of H2G on substantially larger hypergraphs, including a real-world hypergraph with over 90,000 nodes and a synthetic experiment generating over one million hyperlinks (Appendix B.7.3).
  • Generative Fidelity and Domain-Specific Validation [R2W1; R2Q1]

    • We extended our existing domain-specific validation to all three real-world datasets, providing qualitative evidence of plausible patterns (R2W1) (Appendix D.6.3).
    • We further clarified and emphasized that the validity of the generated hyperlinks can be evaluated through our generation fidelity results. These results provide evidence that the generated and observed hyperlinks are close in distribution (R2W1) (Section 4.2).
    • We also optimized our code and allocated more compute budget to H2G-RD tuning. In response to the concern about the performance gap between H2G-EB and H2G-RD on MathOverflow-Answers (R2Q1), H2G-RD now achieves performance more comparable to H2G-EB on this task (Section 4.2).

We thank the reviewers again for their valuable feedback and hope that these revisions and additional results address their concerns.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.