- Rating
- 3(+1)
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 3(+2)
Read review
Paper summary
This paper proposes ADHERE, an unsupervised hypergraph clustering framework designed to mitigate feature, structural, and optimization noise. ADHERE learns stochastic masks over node features and hyperedge incidences to construct feature-denoised and structure-denoised views. It then combines node-level cross-view contrastive learning with clustering-guided self-supervision in a low-dimensional label space. The framework also applies magnitude-based gradient pruning, updating only parameters whose magnitudes exceed a selected percentile. The paper provides perturbation bounds relating feature and propagation-operator differences to representation and clustering errors. ADHERE is evaluated on eight benchmark datasets against graph and hypergraph baselines using ACC, F1, NMI, and ARI. The validation includes component ablations, comparisons with original and randomly perturbed views, mask dynamics, pruning-ratio sensitivity, embedding visualizations, and experiments with alternative hypergraph encoders.
Critical strengths
The research question is of practical significance, the overall pipeline of dual views, representation contrast, and clustering pseudo-labels is relatively complete; the results on the eight datasets are generally competitive; and the recursive proof of Theorem 1 under its strong assumptions is fixable and valid. However, the methodological innovation is mainly a combination of existing techniques, and the current evidence is insufficient to support stronger theoretical and robustness claims. By the way, the figures are beautiful in this paper! I appreciate them!
Critical weaknesses
Both theory and the experiments cannot support the paper's core claim regarding robustness! What a pity!
-
The derivation of Theorem 1 is basically correct, but it cannot prove that ADHERE is effective. Theorem 1 directly assumes and uses δ_X = ‖X^s − X^‖_F and δ_Θ = ‖Θ^s − Θ^‖_2, yet it does not prove at all that the learned masking will reduce them. During training, X^, H^ are not visible, and there is no corresponding constraint in the objective function. Therefore, what Remark 1 calls “the two views complementarily reducing the error upper bound” is only an intuition, not a theorem-derived conclusion.
-
All eight experiments are ordinary clustering benchmarks. There is no controlled injection of feature noise, incorrect incidence, or deletion/addition of hyperedges, nor is there any evaluation of the precision/recall of the mask in identifying known noise. Hyperedges with low label purity are not necessarily noise, and outliers in t-SNE cannot serve as reliable noise ground truth either. Therefore, the experiments at most demonstrate clustering performance on standard data and cannot sufficiently support robust denoising.
Questions for authors
Q1: How do you prove that the learned masks actually reduce and ?
Q2: Can you provide a stability bound connecting to ?
Q3: How exactly is the within-cluster radius in Theorem~2 defined? Why does the proof use instead of ?
Q4: What is the exact distribution of the random variable in Eq. (2)?
Q5: Why are there no controlled experiments with different levels of feature and structural noises?
Q6: Do the curves in the mask analysis represent the masking rate or the retention rate? The descriptions in the main text and appendix appear inconsistent.
Q7: The ``w/o '' variant only removes the regularization term. Why is it described as removing the entire denoising module?
Q8: Why should gradients of small-magnitude parameters be more likely to be noisy? Can you provide evidence based on gradient variance or optimization stability?
8 comments
Thanks for the reminder! We sincerely thank you for your thoughtful evaluation and valuable suggestions, which have helped us further improve our work. Our point-by-point responses are provided below.
Q1: How do you prove that the learned masks actually reduce and ?
R1: Thanks for the comment! To further verify whether the learned denoising masks of our ADHERE can effectively reduce perturbations in practice, we conduct additional experiments on Cora-C by artificially injecting feature and structural noise at varying ratios and compare the corresponding perturbations before and after adaptive denoising. The results are presented below:
| Noise Ratio | 5% | 10% | 20% | 30% |
|---|---|---|---|---|
| Feature Perturbation Ratio (After/Before) | 0.7134 | 0.7028 | 0.6973 | 0.7022 |
| Structural Perturbation Ratio (After/Before) | 0.7933 | 0.8062 | 0.8644 | 0.8865 |
As shown in the results, the learned masks can effectively suppress both feature and structural noise across different noise ratios, i.e., reduce and .
Q2: Can you provide a stability bound connecting to ?
R2: Thanks for the comment! According to Eq. (5), the hypergraph propagation operator can be reformulated as: Correspondingly, the clean hypergraph propagation operator is expressed as . Assume that all node and hyperedge degrees have a common positive lower bound , and the incidence matrices and the fixed diagonal hyperedge weight matrix satisfy: where and are positive bounded constants. By applying the Cauchy--Schwarz inequality, the Lipschitz continuity of over , and the submultiplicativity of matrix norms, we obtain
\leq \left( \frac{1}{d_{\min}} +\frac{B_H(\sqrt{m}+\sqrt{n})}{2d_{\min}^2} \right)\|H-H^\star\|_F.$$ Furthermore, according to the definition of the propagation operator, we have: $$\begin{aligned} \|\Theta-\Theta^\star\|_2 =\|PW_eP^\top-P^\star W_e(P^\star)^\top\|_2 \leq B_W(\|P\|_2+\|P^\star\|_2)\|P-P^\star\|_2 \leq C_H\|H-H^\star\|_F, \quad C_H=\frac{2B_WB_H}{d_{\min}^2}+\frac{B_WB_H^2(\sqrt{m}+\sqrt{n})}{d_{\min}^3}. \end{aligned}$$ The above result establishes a theoretical connection between incidence matrix perturbations and topology-induced propagation perturbations, indicating that under uniform boundedness conditions, reducing incidence matrix perturbations through adaptive structural denoising can further tighten the upper bound on topology-induced propagation perturbations. >Q3: **How exactly is the within-cluster radius $r$ in Theorem~2 defined? Why does the proof use $\sqrt{nr}$ instead of $\sqrt{n},r$ ?** **R3:** Thank you for the comment! In Theorem 2, the within-cluster radius $r$ is defined as the maximum Euclidean distance between a clean node representation and its corresponding clean cluster center, i.e., $$r=\max_{1\leq i\leq n}\|z_i^\star-\mu_{c_i^\star}^\star\|_2.$$ According to the definition of the Frobenius norm, we have: $$\|E^\star\|_F =\sqrt{\sum_{i=1}^{n}\|z_i^\star-\mu_{c_i^\star}^\star\|_2^2} \leq\sqrt{nr^2} =\sqrt{n}r.$$ Therefore, we use $\sqrt{n}r$ in the expression of the upper bound. To avoid ambiguity, **we have provided an explicit definition of the within-cluster radius $r$ in Theorem 2 and clarified the corresponding mathematical expressions in both the theorem and its proof.** >Q4: **What is the exact distribution of the random variable in Eq. (2) ?** **R4:** Thanks for the comment! In Eq. (2), the random variable $\epsilon_{ij}$ follows a standard logistic distribution. Specifically, we first sample $u_{ij}\sim\mathcal{U}(0,1)$ from a uniform distribution and then obtain the random variable through $\epsilon_{ij}=\log u_{ij}-\log(1-u_{ij})$. This reparameterization enables a continuous and differentiable approximation of discrete Bernoulli mask sampling, thereby allowing end-to-end optimization of the denoising masks. **We have added the corresponding definition in Section 2.2 to make it clearer.**Q5: Why are there no controlled experiments with different levels of feature and structural noises?
R5: Thank you for your comment! We additionally inject different proportions of feature and structural noise (ranging from 5% to 30%) into Cora-C and evaluate the two best-performing baselines (HCN [1] and CAHC [2]) alongside our method. The results(ACC)are presented below:
| Method | 5% | 10% | 20% | 30% |
|---|---|---|---|---|
| HCN | 55.63 3.26 | 54.28 4.52 | 47.51 2.86 | 35.47 3.62 |
| CAHC | 65.34 2.76 | 64.28 1.62 | 60.54 1.40 | 54.14 3.28 |
| ADHERE(ours) | 73.01 2.52 | 70.24 1.83 | 65.92 1.45 | 60.57$\pm$ 2.37 |
As shown in the results, the ACC of all methods decreases as the proportion of injected noise increases. However, our method achieves the best performance in most cases and maintains relatively high accuracy across different noise levels, further demonstrating its robustness against artificially injected noise.
[1] A Simple yet Effective Hypergraph Clustering Network, 2025.
[2] From representation to clusters: A contrastive learning approach for attributed hypergraph clustering, 2026.
Q6: Discussion About the Curves in the Mask Analysis.
R6: Thank you for the comment! In our experiments, the curves represent the retention ratios of the feature and structural denoising masks, where higher values indicate that more of the corresponding input information is retained by the model. To avoid ambiguity, we have unified the terminology in Sections 4.5 and E.5 and the corresponding figures, and provided an explicit definition of the retention ratio.
Q7: Discussion About the "w/o " variant.
R7: Thank you for this comment! In the w/o experimental setting, we only remove the denoising regularization term, while the feature and structural denoising masks are still optimized through cross-view contrastive learning and clustering-guided learning. Therefore, this ablation study aims to investigate the role of sparsity regularization in adaptive denoising. In addition, the effectiveness of the entire adaptive denoising mechanism has been validated in Section 4.4 by comparing different view construction strategies. To avoid ambiguity, we have revised the relevant descriptions in Section 4.3 and Appendix E.3 to accurately clarify the setting and purpose of this ablation study.
Q8: Discussion About the Stability of the Parameters Selected by Our Strategy.
R8: Thank you for your question! Upon reviewing our actual implementation, we use the absolute product of a parameter and its gradient to measure parameter importance. This metric provides a first-order approximation of the change in loss induced by removing the corresponding parameter. Selecting parameters for gradient updates based on this metric can reduce interference from optimization signals with limited contributions.
To further investigate the gradient stability of the parameters selected by this strategy, we conduct additional experiments on Cora-C and Mushroom. Specifically, we fix the model parameters and divide them into two groups: high-importance parameters (HIP) and low-importance parameters (LIP) according to our parameter selection strategy. We then repeatedly resample the denoising masks and perform 10 forward and backward passes to calculate the gradient mean and variance of each parameter. For each group, we use the ratio of the sum of gradient variances to the sum of squared gradient means as the evaluation metric (). The results for the two groups are presented below:
| Dataset | Cora-C | Mushroom |
|---|---|---|
| 0.3103 | 0.2236 | |
| 0.0821 | 0.1706 | |
| 3.7791 | 1.3109 |
As shown in the results, LIP consistently exhibits higher than HIP on both datasets, indicating that the gradients of HIP are relatively more stable, whereas those of LIP exhibit greater relative fluctuations and are more susceptible to noise during optimization. These results further support the rationale behind our importance-aware gradient pruning strategy. We have corrected the parameter selection rule in Section 2.5 of the revised manuscript to accurately reflect our original implementation, please refer to it for details.
We are grateful for your time, effort, and thoughtful feedback. We hope that our point-by-point responses, together with the revisions to the manuscript (highlighted in blue), have adequately addressed the issues identified in your review.
ICLR Rebuttal: Please Submit Your Response Soon
This is a reminder that the ICLR rebuttal period is ending soon. I would appreciate it if the authors could submit their response as soon as possible so that I can finalize my review.