acceptodds
Under review as a conference paper at ICLR 2027

CLhypformer: Topology-Aware Hyperbolic Hypergraph Transformer for Graph Representation Learning

Abstract

Learning effective representations from complex graphs requires simultaneously capturing hierarchical geometry and higher-order topology. However, conventional Euclidean embeddings can distort hierarchical relationships, while existing hypergraph construction strategies based on Euclidean proximity or random walks may introduce structurally inconsistent connections. We propose CLhypformer, a topology-aware hyperbolic hypergraph Transformer that jointly addresses these geometric and topological limitations. CLhypformer constructs hyperedges through hyperbolic K-nearest-neighbor relationships, enabling topology construction in a representation space that better accommodates hierarchical structures. It further employs a topology-aware discriminator to generate asymmetric structural views and a debiased three-level contrastive objective to align node, hyperedge, and cross-level representations. Higher-order information is then propagated through a hyperbolic hypergraph Transformer for downstream node classification. Experiments on multiple graph benchmarks show that CLhypformer achieves competitive or superior classification performance and remains robust under severe structural masking. Ablation and qualitative analyses further demonstrate that hyperbolic topology construction and topology-aware contrastive alignment both contribute to the effectiveness of the proposed framework.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Hyperbolic Graph Learning, Hypergraph Neural Networks, Graph Transformers, Contrastive Learning, Topology-Aware Representation Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_oQ5UOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

This paper proposes CLhypformer for node classification on graph data. The model maps node features to a Lorentz space with learnable curvature, constructs hyperedges using hyperbolic KNN, and employs a topology-aware discriminator to generate a structure-preserving view and a strongly perturbed view. It combines node-level, hyperedge-level, and cross-level contrastive objectives with hyperbolic propagation in a joint optimization framework. The authors evaluate the approach through classification, structural masking, ablation, and visualization experiments to support the effectiveness of jointly modeling geometry and topology.

Critical strengths

The paper addresses consistency between the distance used to construct a hypergraph and the space used for representation learning, and attempts to connect this issue with structure-aware contrastive learning. This is a meaningful research question. The experiments combining different construction strategies and contrastive spaces are also relevant to the motivation. However, the available evidence does not establish the independent value of the proposed mechanisms.

Critical weaknesses
  1. The core geometric definitions are inconsistent, and the central learning mechanism is incompletely specified.

Equation (9) does not satisfy the Lorentz-manifold constraint defined in the paper. Interpreting the equation as a node-wise mapping and letting r=∥x∥>0r=\|x\|>0, it gives

\qquad z_s=\frac{\sinh(\sqrt{\kappa}\,r)}{r}x.$$ Consequently, $$\langle z,z\rangle_L=-\frac{1}{\kappa}+\left(1-\frac{1}{\kappa}\right)\sinh^2(\sqrt{\kappa}\,r)$$ This generally differs from the required value of $-1/\kappa$. The spatial component is missing a factor of $1/\sqrt{\kappa}$. Since curvature is learnable, this issue cannot be avoided by assuming unit curvature. For example, with $\kappa=2$ and $\|x\|=1$, the resulting Lorentz inner product is approximately $1.37224$, whereas the constraint requires $-0.5$. If the implementation follows this equation, subsequent distance computations may even fall outside the real-valued domain. Equation (8) also uses inconsistent tangent-space base points: $$\exp_x^c\left(W\log_o^c(x)+b\right).$$ According to Eqs. (6) and (7), $\log_o^c(x)$ lies in the tangent space at the origin, whereas $\exp_x^c$ requires an input in the tangent space at $x$. The paper does not explain how this condition is ensured or provide the necessary parallel transport or projection. **These are questions about whether the geometric operations are valid, not whether the notation is concise.** If they are merely errors in the equations, the authors must specify the operations actually used and explain how they correspond to the experimental implementation. The discriminator's training mechanism is likewise incomplete. Equation (10) produces importance scores, Eqs. (11) and (12) perform discrete Bernoulli sampling, and Eq. (19) includes only classification and contrastive losses. The paper neither provides a separate training objective for the discriminator nor explains how gradients through discrete sampling are handled. Ordinary hard sampling does not directly propagate downstream gradients back to the sampling probabilities. The current description therefore does not establish how the importance scores are learned effectively. Furthermore, the debiasing function, class-prior estimation, and multi-hop topology mask in Eq. (14) are not fully defined. The hyperedge-level and cross-level losses in Eq. (15) are also left unspecified. **The paper presents these mechanisms as central contributions while leaving readers to supply essential parts of the algorithm themselves.** This prevents an adequate assessment of correctness, novelty, and reproducibility. 2. **Hyperbolic graph construction directly overlaps with prior work, while the experiments do not establish the independent contribution of the remaining modules.** Section 2.3 of [Dysformer (a publicly available 2026 preprint)](https://doi.org/10.21203/rs.3.rs-8513633/v1) already describes computing hyperbolic geodesic distances, selecting $K$ nearest neighbors for each node, and constructing an incidence matrix containing $N$ hyperedges. This directly overlaps with the central construction principle in Eqs. (21)–(25). Dysformer also includes learnable curvature and hyperbolic hypergraph propagation. The paper therefore needs to explain the substantive advance of its construction mechanism over this prior work. Using the Lorentz model rather than the Poincaré model does not, by itself, establish a new neighborhood-construction principle: for corresponding points related by an isometry at the same curvature, the two models preserve geodesic distances and nearest-neighbor rankings. Parameterization can affect optimization and numerical behavior, but these differences require justification and evaluation; they do not substitute for an explanation of novelty in graph construction. The complete models are not identical, so CLhypformer may still contribute through topology-aware contrastive learning. However, critical-node-aware augmentation and multi-level contrastive learning already have directly relevant precedents in [CNHCL](https://www.ijcai.org/proceedings/2025/0631.pdf) and [TriCL](https://ojs.aaai.org/index.php/AAAI/article/download/26019/25791). The paper does not sufficiently explain the differences or establish the value of its additions through appropriate comparisons. In particular, CLhypformer− simultaneously removes the discriminator, asymmetric augmentation, and all contrastive objectives. An improvement over this variant only shows that the entire self-supervised branch is helpful. **It does not establish that learned importance scores, asymmetric views, debiasing, or the three contrastive levels are individually effective.** Ordinary random augmentation and contrastive regularization could also explain this improvement. The existing comparisons between construction strategies and contrastive spaces do not rule out this alternative explanation. Moreover, ablations conducted only on Cora do not support claims about module effectiveness across different graph structures. The experiments also present DHG-Bench as a single model, although the cited work is a benchmark containing multiple algorithms. The actual comparator is therefore not clearly identified. The robustness experiments similarly fail to distinguish the effect of topology reconstruction from the contribution of the contrastive mechanism. If the proposed model can reconstruct connections from features while a baseline must rely on a corrupted original topology, the difference between their performance curves cannot be directly attributed to the proposed robustness mechanism. **The current experiments support the performance of the complete system, but not the claimed contributions of its individual designs or the mechanisms offered to explain those results.**
Questions for authors
  1. Do the geometric operations match the actual implementation?

Please explain how Eq. (9) satisfies the Lorentz constraint and how Eq. (8) handles tangent spaces with different base points. If the equations differ from the code, please provide the actual operations and report manifold-constraint residuals and checks of the validity of distance computations under learnable curvature.

  1. How are the topology-aware discriminator and debiased objective trained?

Please specify the discriminator's optimization objective, parameter-update procedure, and treatment of gradients through the discrete sampling in Eqs. (11) and (12). Please fully define Debias\mathrm{Debias}, class-prior estimation, MtopoM_{\mathrm{topo}}, Ledge\mathcal{L}_{\mathrm{edge}}, and Lcross\mathcal{L}_{\mathrm{cross}}. Without these definitions, it is not possible to determine whether the implementation supports the stated contributions.

  1. What minimal ablations establish the independent value of the added modules?

Please provide at least the following comparisons using the same backbone, comparable training budgets, and consistent loss normalization:

  • The gains exceed those of ordinary contrastive regularization: Compare supervised training only, random augmentation with standard contrastive learning, and the full model. Keep the contrastive levels and training conditions comparable.

  • The importance scores contain useful structural information: Compare learned WcritW_{\mathrm{crit}}, scores shuffled among existing associations, and uniform random association removal. Preserve or match the perturbation budget for each hyperedge to avoid attributing differences in removal rates to importance learning.

  • Asymmetric views are necessary: Compare preservation/preservation, preservation/strong perturbation, and random/random view pairs, with matched actual average removal rates for the respective views.

  • The topology mask and debiasing each contribute: Compare standard contrastive learning, topology masking only, debiasing only, and both components together, with all other components fixed.

  • The three contrastive levels are complementary: Compare node only, node plus hyperedge, node plus cross-level, and all three objectives, with controlled loss scales and tuning budgets.

These core ablations should cover at least three datasets with different structural characteristics, using consistent splits and repeated runs. The relevant question is whether removing a specific mechanism consistently removes its claimed benefit, rather than merely whether removing the entire self-supervised branch reduces performance.

  1. Does the advantage of hyperbolic graph construction depend on a particular neighborhood size K?

Please analyze sensitivity to KK in Eqs. (22) and (23), for example using K∈{5,10,20,40,80}K\in\{5,10,20,40,80\}, and include the actual default value.

Compare hyperbolic KNN and Euclidean KNN simultaneously under the same KK, backbone, and tuning budget, and report classification performance, variability, and construction cost.

This experiment should establish whether hyperbolic construction remains effective across a reasonable range of neighborhood sizes or whether its advantage is confined to one particular setting. A curve showing only the full model's performance as KK changes does not establish the independent benefit of the geometric choice.

  1. Does the proposed contrastive mechanism outperform ordinary contrastive learning across a reasonable range of loss weights?

Please analyze sensitivity to λ\lambda in Eq. (19), for example using λ∈{0,10−3,10−2,10−1,1,10}\lambda\in\{0,10^{-3},10^{-2},10^{-1},1,10\}, and include the actual default value.

With loss normalization, graph construction, and other settings held fixed, compare the proposed contrastive branch against a branch using ordinary random augmentation and standard contrastive learning. Use the same validation-based tuning budget and report the relative magnitudes of the supervised and weighted contrastive terms.

The key question is not whether a positive contrastive loss weight outperforms a zero weight, but whether the proposed design outperforms ordinary contrastive learning under comparable tuning conditions. If a standard contrastive objective achieves the same performance with equal tuning, the necessity of the additional mechanisms remains unestablished.

  1. Do the technical contribution and robustness gains hold against directly relevant alternatives?

Please clarify the methodological differences from Dysformer, CNHCL, and TriCL, and provide direct comparisons under consistent data splits, comparable supervision, and training budgets. Please also identify the specific algorithm represented by “DHG-Bench.”

For structural masking, please specify whether masking is applied to the original graph, the reconstructed hypergraph, or the augmented views, and whether all methods have the same opportunity to reconstruct topology. Comparing the full model, CLhypformer−, and an ordinary contrastive-learning baseline under the same reconstruction conditions is necessary to determine whether the robustness gains arise from the proposed mechanism.

Reviewer_YQ9XOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CLhypformer, a topology-aware hyperbolic hypergraph Transformer for node classification. It maps node features to a Lorentz manifold with learnable curvature and constructs hyperedges using hyperbolic nearest neighbors. A topology-aware discriminator scores node–hyperedge associations to generate two asymmetric views: one preferentially preserves high-scoring associations, while the other preferentially perturbs them. Training combines a three-level contrastive objective with supervised node classification. The experiments cover six datasets, with additional studies on structural masking, geometry/topology ablations, and qualitative visualizations.

Critical strengths
  1. The paper studies a relevant question: how the geometry used to learn representations should relate to the construction of higher-order neighborhoods. Using the same geometry for both is a reasonable approach for hierarchical data.
  2. The two views provide an interesting alternative to uniform deletion. One tends to preserve high-scoring associations, while the other targets them for perturbation. If the scores reflect structural importance, this could help the model learn from perturbations while keeping an informative reference view.
  3. The experiments include a supervised-only variant with the same geometric backbone, as well as comparisons of hyperbolic KNN, Euclidean KNN, and random-walk construction. The masking experiments and visualizations also provide some insight beyond the main accuracy results.
Critical weaknesses
  1. Some geometric formulas are inconsistent. Under the manifold definition in Eq. (1), the spatial component in Eq. (9) is missing a factor of (1/\sqrt{\kappa}). Eq. (8) uses different base points for the logarithmic and exponential maps without explaining the mapping between tangent spaces. These may be writing errors, but the authors need to clarify which formulas were used in the implementation.
  2. Key parts of the method are not fully defined. How the criticality discriminator is trained, and how gradients are handled through Bernoulli sampling, are not explained. The topology mask, Debias term, and hyperedge and cross contrastive losses also need clear definitions. Without these details, it is difficult to check or reproduce the method.
  3. The novelty and the role of each component need to be clearer. TriCL already uses multi-level hypergraph contrast, while Critical Node-aware Augmentation uses a discriminator to preserve important structure. The paper should explain how it differs from these methods and compare against them directly. Removing all contrastive components at once does not show how much each proposed component contributes.
  4. The topology and robustness claims need stronger evidence. Using hyperbolic KNN does not by itself show that the resulting neighborhoods better preserve hierarchy. Figure 6 compares different runs or checkpoints, making the claimed topology improvement harder to assess. For Figure 3, the paper should explain what is masked, whether KNN is rebuilt afterward, and whether the methods receive equivalent corrupted inputs.
Questions for authors
  1. Are Eqs. (8) and (9) writing errors? What exponential map and aggregation rule were used in the implementation? Please explain how the manifold constraints are maintained as curvature changes and whether the experiments use the corrected formulas.
  2. How is the criticality discriminator trained when the masks are sampled from Bernoulli distributions? Please give its training objective and explain how gradients are computed. Please also define the topology mask, debias term, hyperedge loss, and cross loss, including how the class prior is estimated without using held-out labels.
  3. What are the main differences from TriCL and Critical Node-aware Augmentation in the loss and discriminator design? With the backbone and training budget held fixed, do asymmetric views and debiasing improve over uniform augmentation and the loss without debiasing?
  4. Is the original adjacency used alongside the feature-based hypergraph, and when is KNN recomputed? What is masked in the robustness experiments, and do all methods receive equivalent corrupted information? Can Figure 6 use matched runs and checkpoints, a common normalization, and a quantitative measure of topology quality?
Reviewer_cJ2bOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

CLhypformer combines Lorentz-space KNN hypergraph construction, a hyperbolic hypergraph Transformer, and topology-aware contrastive learning for node classification. A criticality discriminator generates structure-preserving and hard-perturbed views, while supervised classification is combined with multilevel debiased alignment. The paper reports favorable accuracy on six benchmarks, robustness to structural masking, and gains over a variant without self-supervision.

Critical strengths
  1. Coherent geometry-topology motivation. Constructing higher-order neighborhoods in the representation geometry is a reasonable goal. The center-plus-K-neighbors construction in §3.2 is explicit, and Eqs. (11)-(12) explain the complementary retention rules underlying the proposed augmentation.
  2. Useful controls and stress tests. CLhypformer- preserves hyperbolic KNN and the backbone (§3.3), providing a control for the complete self-supervised package. This controlled comparison helps separate its overall contribution from changes in representation geometry or backbone capacity. Figure 4 compares construction/contrastive-space combinations, and Figure 3 evaluates progressively stronger structural masking.
Critical weaknesses
  1. The central method needs consistent and complete specification. Equation (9), as written, omits the inverse square root curvature factor in its spatial component and therefore does not satisfy the Lorentz constraint in Eq. (1) for general learnable curvature. This affects the geometric basis of the distance and KNN calculations in Eqs. (21)-(22), although it may be a typographical error rather than the implemented mapping. Moreover, Eq. (14) leaves Debias and the exact reachability mask undefined, Eq. (15) omits the hyperedge and cross-level loss definitions, and Eqs. (10)-(12) do not explain discriminator training through Bernoulli sampling. These gaps prevent reproduction of the core contribution. Please provide the mapping actually used and a complete algorithm defining the losses, criticality-training objective, and gradient estimator or alternative optimization procedure.
  2. The specific topology-aware contribution remains insufficiently established. CLhypformer- removes the discriminator, asymmetric views, and all contrastive objectives together (§3.3). Figure 4 therefore supports the package as a whole, without distinguishing topology-aware augmentation or false-negative correction from ordinary contrastive regularization. The geometry/construction ablations (§4.3) do not resolve this distinction. Supplementary Tables 1-3, cited in §§4.1-4.3, are also absent from the supplied 11-page PDF; data splits, tuning budgets, run counts, and masking before versus after KNN reconstruction are unspecified. This limits assessment of comparison fairness and the robustness mechanism. Please provide the protocol and numerical results with uncertainty, and compare uniform augmentation and a non-debiased contrastive objective under matched backbone, geometry, input information, and expected deletion rates.
Questions for authors
  1. How is the topology-aware discriminator trained through the Bernoulli masks in Eqs. (11)-(12)?
  2. Under matched information and perturbation budgets, what evidence distinguishes topology-aware augmentation and debiasing from generic contrastive regularization?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.