acceptodds
Under review as a conference paper at ICLR 2027

HHGD: Hyperbolic Hypergraph Diffusion for Valid Hypergraph Generation

Abstract

Hypergraphs provide a natural representation for higher-order relations in social interactions, biological systems, recommender systems, and scientific networks, where pairwise graphs cannot faithfully capture group structure. Generating such data requires learning native node–hyperedge incidence patterns while producing discrete hyperedges with realistic cardinalities and non-degenerate overlaps. We propose HHGD, a hyperbolic latent diffusion framework that jointly represents nodes and hyperedges as tokens in hyperbolic space. HHGD combines an incidence-aware hyperbolic encoder, numerically stable diffusion in the origin tangent space, a curvature-aware hyperbolic score network, and a cardinality-aware Poincar\'e-distance decoder that maps generated tokens to binary incidence matrices. Experiments on synthetic and real-world hypergraph datasets show that HHGD can produce samples with exact uniqueness and novelty, while strict validity varies substantially across datasets and is strongest on several synthetic regimes, attaining the highest displayed mean strict validity on ER, Hypergraph SBM, Tree, and Tags-Math while remaining competitive in matching local and global structural distributions. Comprehensive architecture ablations further show that the hyperbolic latent components provide a coherent inductive bias for higher-order incidence generation, with their strongest benefits appearing in validity and hyperedge-size fidelity, while improvements in global topology remain dataset dependent. These results demonstrate the promise of hyperbolic latent diffusion for valid native hypergraph generation and clarify where its structural benefits are strongest.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Hypergraph Generation, Diffusion Models, Hyperbolic Representation Learning, Geometric Deep Learning, Graph and Hypergraph Neural Networks

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_NAL7OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes HHGD, a latent diffusion model for hypergraph generation. An encoder maps the padded incidence matrix to node and hyperedge tokens in the Poincaré ball, using tangent-space message passing with Lorentz re-embeddings and a Transformer. A DDPM is then trained on the origin-tangent coordinates of these tokens with a Poincaré-space denoiser. A decoder converts the generated tokens into a binary incidence matrix by ranking Poincaré distances and selecting the top-kk nodes per hyperedge, where kk comes from a cardinality head rescaled by a factor γ\gamma chosen by grid search on validation data. The method is evaluated on four synthetic and four real-world datasets.

Critical strengths

Generating hypergraphs directly from the incidence structure, with explicit modeling of hyperedge cardinalities, is a relevant problem. HHGD obtains 100% strict validity on the synthetic ER, SBM, and Tree datasets.

Critical weaknesses

I recommend rejection for this paper. The method looks like a direct application of hyperbolic graph diffusion to the incidence matrix, and it relies on a long list of ad-hoc components rather than on a principled design.

  • Limited novelty. HHGD is essentially hyperbolic latent graph diffusion (Fu et al., 2024) applied to the incidence matrix. The contrast in Table 1 between "bipartite" (HYGENE) and "native incidence" (HHGD) representations is not meaningful: the incidence matrix HH is exactly the biadjacency matrix of the bipartite graph, i.e., the adjacency matrix of the bipartite graph is the block matrix with zero diagonal blocks and HH and H⊤H^\top as off-diagonal blocks. The hyperbolic components also add little. In the encoder, the exponential map at the end of each layer is undone by the logarithmic map at the start of the next one, and the denoiser uses a constant context CC (all entries equal to 1/M1/M, Eqn. (D.7)) that carries no incidence information. Consistent with this, HHGD and Euc+Hyp are almost identical on every dataset (e.g., validity 1.7 vs. 2.7, 31.7 vs. 31.3, 36.0 vs. 36.3, and 89.0 vs. 89.0 on the real-world datasets).

  • The method stacks ad-hoc fixes without justification or ablation. Several components look like patches added one by one to fix specific problems:

    • the positive-class weight of 9 in the BCE;
    • the topology loss in Eqn. (E.8) and the boundary loss;
    • clipping the predicted clean-token norms to 3.0 at every sampling step;
    • the hyperedge sizes, which come from a threshold scale γ\gamma selected by grid search on validation data using the same family of structural metrics reported in the tables.

    None of these choices is motivated or ablated. As a result, the validity is largely enforced by the decoder (it removes hyperedges of size smaller than two and suppresses duplicates) rather than learned by the generative model. The baselines don't get the same post-processing, so the validity comparison isn't fair. Even with these heuristics, validity is low on several datasets (1.7% on Contact and 5.0% on Ego).

  • Permutation invariance, scalability, and expressivity. The hyperedge slots are ordered lexicographically under the native node labeling. Contrary to Appendix E.4, this is not a canonicalization: it doesn't address the permutation symmetry of hypergraphs, and the model is not permutation invariant, as the paper acknowledges in Section 3.1. The node and hyperedge budgets NN and MM must be chosen in advance, the cost is O((N+M)2d)O((N+M)^2 d) per Transformer layer, and the largest dataset has only 130 nodes, so it's unclear whether the method scales. I'm also concerned about the expressivity of the decoder, since all hyperedge sizes are set by a predicted fraction of NN rescaled by a single global γ\gamma.

Other comments:

  • The method section is hard to follow. It mixes the method with implementation details (e.g., "scales KℓK_\ell are converted to Python scalars"), and Phase II is mentioned in Section 3.4 before Phase I is defined in Section 3.6.
  • Uniqueness and novelty of 100% are not informative, since almost all baselines also reach these values.
  • Typo: "projection.The" (line 94).
Questions for authors

Clarifications:

  • Can the decoder generate hypergraphs with only small hyperedges, or with hyperedge-size distributions different from the training set, given that the sizes are set by a predicted fraction of NN and a single global γ\gamma? How is the number of active hyperedges determined?
  • Why is the norm clipping during sampling needed, and what happens without it?
  • What is the validity of the baselines if the same decoding rules are applied (size at least two, duplicate removal, calibrated top-kk)?
  • How were the baselines tuned? Some HYGENE results are surprising, e.g., OOM on Email-Eu, which has only 64 node slots and 128 hyperedge slots, and 14.3% validity on Ego.

Suggestion for the revision:

  • Ablate the ad-hoc components (positive-class weight, topology and boundary losses, norm clipping, and the γ\gamma calibration) and compare against a fully Euclidean pipeline with the same decoder.
Reviewer_Pz63OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes HHGD, a two-stage latent diffusion model for hypergraph generation. Stage I trains an incidence autoencoder: a padded incidence matrix H∈{0,1}N×MH\in\{0,1\}^{N\times M} is encoded by HGCN-style tangent-space message passing plus a Euclidean Transformer into Poincaré-ball tokens for nodes and hyperedges; a decoder scores node–hyperedge membership via normalized Poincaré distance, predicts each hyperedge slot’s cardinality, and discretizes via validation-calibrated per-edge top-kk selection. Stage II freezes the autoencoder, maps tokens to the tangent space at the origin with log⁡0\log_0, and trains a standard DDPM there; the denoiser uses Poincaré-space attention/feed-forward blocks and a fixed uniform context C=M−111⊤C=M^{-1}\mathbf 1\mathbf 1^\top. Sampling uses 50-step DDIM. Experiments cover four synthetic (ER, hypergraph SBM, Tree, Ego) and four real temporal datasets (Contact, NDC, Email-Eu, Tags-Math), each with 350/50/100 hypergraphs, against HyperPA, HyperLap+, HYGENE, and VAE/GAN/incidence-diffusion adaptations, plus a 2×2 encoder × score-network ablation. HHGD reaches 100% strict validity on ER/SBM/Tree and 89% on Tags-Math, with uniqueness and novelty at 100%, but validity on Contact (1.7%), Ego (5.0%), NDC (31.7%), and Email-Eu (36.0%) is low, and structural-distance results are mixed.

Critical strengths

Native incidence modeling. Representing nodes and hyperedges jointly as tokens and decoding via node–hyperedge distances avoids clique expansion and pairwise surrogates, which is a sensible design for hypergraphs.

Cardinality-aware decoding is a reasonable mechanism for producing discrete hyperedges with controlled sizes, and the duplicate/singleton suppression is clearly specified.

Critical weaknesses
  1. The “hyperbolic” contribution is largely nominal and never isolated. Diffusion is plain Euclidean DDPM in the tangent space at the origin, where exp⁡0/log⁡0\exp_0/\log_0 reduce to a radial tanh⁡\tanh reparameterization; Appendix C.6 concedes no manifold density is modeled. Curvatures are fixed constants with no gradients. The encoder is tangent-space message passing followed by a Euclidean Transformer. The ablation does not isolate geometry either: the encoder swap (HGCN vs. Euclidean) changes almost nothing on any dataset (Tables 3, 7) — Euc+Hyp is actually best on validation (App. F.5) — and the score-network swap compares the proposed denoiser to a “Lightweight” network that is a different architecture (App. D.5: “not obtained merely by deleting a single Lorentz map”), confounding geometry with capacity. The decoder — the only component where Poincaré distance actually enters the output — is never ablated against a Euclidean-distance decoder. The abstract’s claim that hyperbolic components “provide a coherent inductive bias” is therefore unsupported by the paper’s own experiments.

  2. Strict validity is largely a decoder artifact, and the title’s claim is not borne out. Validity requires the correct active-node count, no empty/singleton/duplicate hyperedges. On the fixed-size synthetic datasets (no padding, all nodes active), top-kk decoding with k≥2k\ge2 and duplicate suppression satisfies these conditions by construction, so 100% validity on ER/SBM/Tree is essentially guaranteed. On padded real data, where the model must also produce the right active-node count, validity collapses (1.7% on Contact, 5.0% on Ego). A method titled “for Valid Hypergraph Generation” that is valid on 2–36% of real samples does not support its title. Conversely, 0% validity for HyperPA on ER (a well-established generator) suggests the validity criterion mainly penalizes node-count mismatches, which has little to do with generation quality.

  3. Structural fidelity is frequently worse than simple baselines. HyperLap+ beats HHGD on nearly every structural metric on ER, SBM, Tree, and Tags-Math, often by large margins (Tree: EdgeCnt 0.14 vs. 1.80; NodeDeg 0.487 vs. 0.601; Spectral 0.053 vs. 0.083). On Contact, HHGD has the worst spectral MMD of all seven methods (0.510) and EdgeSize W1 of 0.931 vs. 0.049 for HyperPA — directly contradicting the claimed “hyperedge-size fidelity.” On NDC, HyperPA reaches 93% validity vs. HHGD’s 31.7%. The abstract’s “remaining competitive in matching local and global structural distributions” is overstated.

  4. The denoiser’s “incidence-aware” blocks are degenerate. With C=M−11N1M⊤C=M^{-1}\mathbf 1_N\mathbf 1_M^\top at both training and sampling, the hypergraph convolution in the denoiser (App. D.2) reduces to uniform mean pooling over all tokens — no incidence information is used. The paper flags this as a limitation but still lists an “incidence-aware ... score network” as a contribution.

Questions for authors
  1. What exactly is the “Lightweight” score network — is it Euclidean, and is it parameter-matched? Please add a parameter-matched Euclidean denoiser (and a Euclidean-distance decoder) so the effect of hyperbolic geometry can actually be isolated.

  2. For each dataset, please break down validity failures by cause (wrong active-node count, singleton edge, duplicate edge, empty sample). Is 100% validity on fixed-size synthetic data achievable by any top-kk decoder with k≥2k\ge2 and dedup, regardless of the generative model?

  3. Why does HyperPA score 0% validity on ER and HyperLap+ 0% on Contact? Which condition fails? Do baselines that generate variable node counts have a fair chance under the “exact active-node count” requirement?

  4. With the uniform context CC, does the denoiser’s hypergraph block reduce to global mean pooling? If so, what distinguishes it from a DiT with mean-pooling layers?

  5. Was any calibration analogous to γ\gamma selection applied to baselines (e.g., thresholding for Inc. Diff.)? How sensitive are HHGD’s structural metrics to γ\gamma?

  6. How does the model infer the active-node count on padded real datasets when no masks are used in the encoder, Transformer, or losses? Is this the primary driver of low real-data validity?

Reviewer_BWZWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes a new diffusion model to generate hypergraphs while retaining validity constraints. The model relies on a latent representation based on hyperbolic geometry. The model retains the validity constraints set by the authors despite losing in overall generative quality.

Critical strengths

The submitted manuscript addresses the problem of hypergraph generation. This is a topic of growing interest in the scientific community that poses complex challenges.

Critical weaknesses

Major

  1. Theoretical justification The authors mention hyperbolic geometry being widely used for embedding real-world networks. While this holds true for some types of networks, I cannot recall any study of this kind for hypergraphs, and the authors do not seem to mention any in their references.

    a. On network geometricity I do not see what is the rationale behind using geometric embedding for non-geometric data. I do not understand how -for instance- Erdos Renyi hypergraphs, which entirely lack geometricity can gain something out of this embedding strategy.

    b. On the connection between hyperbolic space and validity It is not clear to me how the methodology proposed in the paper can improve validity rate of the generated hypergraphs

  2. Different validity definition The definition of validity used in the paper is different from the ones used in the prior art, where -for instance- a generated tree graph is considered valid if it is acyclic. The used definition is different in the submitted manuscript. I suggest changing the used term to avoid confusion in the reader. I further encourage the authors to provide the validity measures of their generated graphs, to make an easier and fair comparison with the prior work.

    a. On the baseline performance Indeed, the validity rate of the baseline HYGENE is orders of magnitude lower than the one reported in their paper and other reproductions (e.g., 77% -> 0% for Tree hypergraphs)

Less critical

  1. Metrics Most works on hypergraph and graph diffusion models use MMD (maximum mean discrepancy) to assess the quality of generated data. I do not see any issue with using W1 in terms of soundness. However, using a different metric than the de-facto standard can introduce complications in using the model as a baseline for future works in the field. I got confused with graph diffusion models works.
  2. Similarity with state-of-the-art Several parallelism could be drawn between this work and the one by Fu et al. [1] in which their graph diffusion model runs on a hyperbolic latent representation.

[1] Fu, Xingcheng, et al. "Hyperbolic geometric latent diffusion model for graph generation." arXiv preprint arXiv:2405.03188 (2024).

Questions for authors

I pose the following questions to the authors

  1. What is the motivation behind generating hypergraphs with validity as defined in appendix F.6?
  2. Could the authors please provide some theoretical justification behind the use of hyperbolic geometry for embedding hypergraphs?
  3. How does the use of hyperbolic geometry connect to the validity of the generated hypergraphs? Especially, how does this connect with the validity of the generated hypergraphs that are not possible to embed in geometric spaces (e.g., Erdos Renyi and SBM)?
  4. Could the authors please provide the results in validity as specified in most state-of-the-art papers? Consider, for instance the one used in HYGENE
  5. I strongly encourage the authors to rephrase the term validity with a more appropriate one to avoid confusion in the reader.
  6. Could the authors please provide the generative quality results measured as MMDs?
  7. Could the authors please describe what are the key differences between their work and the one by Fu et al.? (see weaknesses)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.