- 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- nodes per hyperedge, where comes from a cardinality head rescaled by a factor 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 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 and 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 (all entries equal to , 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 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 and must be chosen in advance, the cost is 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 rescaled by a single global .
Other comments:
- The method section is hard to follow. It mixes the method with implementation details (e.g., "scales 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 and a single global ? 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-)?
- 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 calibration) and compare against a fully Euclidean pipeline with the same decoder.