acceptodds
Under review as a conference paper at ICLR 2027

ADHERE: Adaptive Denoising for Robust Hypergraph Clustering

Abstract

Hypergraph clustering aims to discover cluster structures by jointly exploiting node features and high-order relationships encoded by hyperedges. However, real-world hypergraphs inevitably contain feature and structure noise, which can propagate throughout the clustering process by degrading input reliability, distorting learned representations, and introducing unstable optimization signals. To address these issues, we propose an adaptive denoising-aware joint optimization framework for hypergraph clustering that tackles noise from input, representation, and optimization perspectives. Specifically, we design a differentiable denoising mechanism based on the reparameterization trick to adaptively select reliable node features and hyperedge incidences, thus generating two complementary denoised views. Then we perform cross-view contrastive learning in a high-dimensional space to learn consistent representations, while introducing pseudo-label-based self-supervised learning in a low-dimensional label space to explicitly reinforce cluster-level semantic structures. Furthermore, we adopt a parameter-pruning gradient update strategy to suppress unstable optimization signals during training. Extensive experiments on multiple benchmark datasets demonstrate the effectiveness and robustness of the proposed framework across several clustering metrics.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Neural Networks, Hypergraph Clustering, Denoising Learning

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.40 over 5 reviews, above 99% of the others.
Reviewer_YKdHOpenReview ↗
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!

  1. 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.

  2. 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 δX\delta_X and δΘ\delta_\Theta?

Q2: Can you provide a stability bound connecting ∥H−H⋆∥\|H-H^\star\| to ∥Θ−Θ⋆∥\|\Theta-\Theta^\star\|?

Q3: How exactly is the within-cluster radius rr in Theorem~2 defined? Why does the proof use nr\sqrt{nr} instead of n r\sqrt{n}\,r?

Q4: What is the exact distribution of the random variable ϵ\epsilon 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 Lden\mathcal{L}_{\mathrm{den}}'' 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
Reviewer_YKdHOpenReview ↗

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.

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 δX\delta_X and δΘ\delta_{\Theta}?

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 Ratio5%10%20%30%
Feature Perturbation Ratio (After/Before)0.71340.70280.69730.7022
Structural Perturbation Ratio (After/Before)0.79330.80620.86440.8865

As shown in the results, the learned masks can effectively suppress both feature and structural noise across different noise ratios, i.e., reduce δX\delta_X and δΘ\delta_{\Theta}.

Q2: Can you provide a stability bound connecting ∣H−H∗∣|H-H^* | to ∣Θ−Θ∗∣|\Theta-\Theta^*| ?

R2: Thanks for the comment! According to Eq. (5), the hypergraph propagation operator can be reformulated as: Θ=PWeP⊤,P=Dv−12HDe−12.\Theta=PW_eP^\top,P=D_v^{-\frac{1}{2}}HD_e^{-\frac{1}{2}}. Correspondingly, the clean hypergraph propagation operator is expressed as Θ⋆=P⋆We(P⋆)⊤\Theta^\star=P^\star W_e(P^\star)^\top. Assume that all node and hyperedge degrees have a common positive lower bound dmin⁡>0d_{\min}>0, and the incidence matrices and the fixed diagonal hyperedge weight matrix satisfy: ∥H∥2,∥H⋆∥2≤BH,∥We∥2≤BW,\|H\|_2,\|H^\star\|_2 \leq B_H,\|W_e\|_2\leq B_W, where BHB_H and BWB_W are positive bounded constants. By applying the Cauchy--Schwarz inequality, the Lipschitz continuity of f(x)=x−1/2f(x)=x^{-1/2} over [dmin⁡,+∞)[d_{\min},+\infty), 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:

Method5%10%20%30%
HCN55.63 ±\pm 3.2654.28 ±\pm 4.5247.51 ±\pm 2.8635.47 ±\pm 3.62
CAHC65.34 ±\pm 2.7664.28 ±\pm 1.6260.54 ±\pm 1.4054.14 ±\pm 3.28
ADHERE(ours)73.01 ±\pm 2.5270.24 ±\pm 1.8365.92 ±\pm 1.4560.57$\pm$ 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 Lden\mathcal{L}_{den}" variant.

R7: Thank you for this comment! In the w/o Lden\mathcal{L}_{den} 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 (RgR_g). The results for the two groups are presented below:

DatasetCora-CMushroom
RgLIPR_g^{LIP}0.31030.2236
RgHIPR_g^{HIP}0.08210.1706
RgLIP/RgHIPR_g^{LIP} / R_g^{HIP}3.77911.3109

As shown in the results, LIP consistently exhibits higher RgR_g 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.

Reviewer_YT1SOpenReview ↗
Rating
4
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the problem of hypergraph clustering under feature and structural noise and proposes an adaptive denoising framework named ADHERE. It first employs learnable stochastic masks to denoise the hypergraph and constructs two complementary views, then it adopts cross-view contrastive learning and cluster-guided learning. Additionally, a gradient pruning strategy is introduced to improve training stability during the optimization process. Extensive experiments demonstrate that ADHERE achieves competitive results.

Critical strengths
  1. The motivation is clear. The paper identifies three key limitations of existing methods concerning input reliability, clustering-related semantic guidance and optimization stability, and proposes corresponding modules to address these limitations.

  2. The denoising strategy is well-designed. The proposed method jointly addresses feature and structural noise, providing an insightful denoising strategy for hypergraph clustering.

Critical weaknesses
  1. Although the framework learns node representations from two denoised views, the paper does not specify how the final representations used for K-means clustering are obtained from these views.

  2. In the ablation study, w/o LdenL_{\mathrm{den}} is described as removing the denoising module. However, removing the denoising regularization loss may not be equivalent to removing the entire denoising mechanism. The authors need to provide a more explicit clarification of the precise ablation setting and its corresponding conclusions.

Questions for authors
  1. How are the final node representations for K-means clustering constructed from the two denoised views?

  2. In section 4.5 and E.5, masking rate and retention ratio are used in the analysis, do they represent the same meaning? Please provide a consistent definition and clarify what the corresponding experimental curves represent.

3 comments

We greatly appreciate your time and effort in evaluating our manuscript, and we address each of your concerns in the following responses.

Q1: About the Final Node Representations.

R1: Thank you for your question! During training, we obtain node representations from the two complementary denoised views through a shared hypergraph encoder and enhance their semantic consistency and clustering discriminability through cross-view contrastive learning and clustering-guided learning. During inference, we average the node representations from the two views to obtain the final node representations, thereby integrating their complementary feature and structural semantic information. We then perform K-means clustering on the final node representations to obtain the clustering results. Accordingly, we have added the corresponding implementation details in Section E.3 to make this process clearer.

Q2: About the Retention Ratio.

R2: Thank you for your comment! In our experiments, the corresponding 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 Section 4.5 and the corresponding figures and provided an explicit definition of the retention ratio.

Q3: About the Ablation Variant "w/o Lden\mathcal{L}_{den}.

R3: Thank you for your comment! In the "w/o Lden\mathcal{L}_{den}" 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. The experimental results demonstrate that sparsity regularization promotes the selective retention of input information, thereby improving clustering performance. In addition, in Section 4.4, we further validate the effectiveness of the entire adaptive denoising mechanism by comparing the adaptively denoised views with the original views without denoising. We have revised the relevant descriptions in Section 4.3 and Section E.3 to clarify the precise experimental setting and corresponding conclusions of this ablation study.

We sincerely appreciate your time and consideration. We hope the revised manuscript (changes are highlighted in blue) provides a clearer and more comprehensive presentation of our work.

My concerns have been addressed.

Thanks for your valuable feedback, and we're glad to see that our responses have addressed your concerns.

Reviewer_GJXdOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This manuscript studies the noise problem in node-level hypergraph clustering, noting that real-world hypergraphs contain both feature noise and structure noise (low-purity hyperedges account for 22.34% in the illustrated data), while existing contrastive hypergraph clustering methods based on random augmentation ignore noise and are affected by it at the input, representation, and optimization levels. The authors propose the ADHERE framework, whose core is a differentiable denoising module based on the reparameterization trick: Bernoulli masks are learned for node feature entries and hyperedge incidences, and after continuous relaxation yield a feature-denoised view and a structure-denoised view, with a sparsity regularizer preventing the masks from degenerating to keeping everything. The two denoised views are passed through a shared HGNN encoder, followed by node-level cross-view contrastive learning in a high-dimensional space and clustering-guided self-supervised learning with k-means pseudo-labels in a low-dimensional label space, with the three losses optimized jointly.

Critical strengths
  1. Clear motivation with a well-grounded starting point. The paper splits the noise in hypergraph clustering into feature noise and structure noise, and backs this up with quantitative evidence (Figure 1).

  2. The method design and writing are clear. The three levels of the design, input, representation, and optimization, follow one coherent storyline.

  3. The paper comes with a theoretical analysis that matches the method design. Theorem 1 and Theorem 2 line up with the two denoised views and the clustering objective, respectively.

  4. The experiments cover a lot of ground. Eight datasets, eight baselines, ablations on every loss term and view, plus analyses of view construction, denoising masks, hyper-parameters, visualization, and backbone generalization.

Critical weaknesses
  1. The related work misses the closest line of literature. Learnable Bernoulli masks with the reparameterization trick already have a series of works in graph learning, such as AD-GCL, GSAT, and hard-concrete structure learning; cross-view contrastive learning combined with k-means pseudo-labels aligned by the Hungarian algorithm for clustering guidance is also common practice in multi-view clustering and in hypergraph clustering methods such as CAHC; and gradient pruning is a standalone optimization trick. The combination of the three is new, but none of the individual components is.

  2. There is no bridge between the theoretical analysis and the claim that denoising is effective. Theorem 1 is a standard layer-wise Lipschitz propagation perturbation bound that holds for any perturbation and does not depend on whether the masks are learned well. Theorem 2 is a standard K-means stability bound whose assumptions (K-clusterable, number of misclustered nodes smaller than the smallest cluster size) cannot be verified. Both theorems could be applied as-is to any two-view hypergraph contrastive method.

  3. The margins in the main table are small. ADHERE is best in 21 of the 32 dataset–metric combinations. On Cora-A it falls behind CAHC on all four metrics, by 1.3 to 4.2 points.

  4. The definition of the ablation variants is ambiguous. "w/o Lden" is described in the appendix as "removing the whole denoising module", but according to the equations it seems to only remove the sparsity regularizer, while the masks are still being learned.

Questions for authors

Please refer to Weaknesses。

5 comments

We sincerely thank you for your thorough review and constructive feedback. We have carefully considered your comments and provide detailed responses below.

Q1: Discussion About the Related Work and Methodological Differences.

R1: Thank you for your valuable comment! The methodological differences of our ADHERE primarily lie in the design of its adaptive denoising mechanism and joint optimization strategy. While most existing methods employing learnable Bernoulli masks perform denoising within the same view, we propose a decoupled and complementary adaptive denoising mechanism that constructs two complementary views for feature and structural denoising. This design is motivated by the fact that in unsupervised hypergraph clustering, learned masks may mistakenly filter out informative content, while simultaneously denoising features and structures may further exacerbate information loss. By constructing two separate denoised views, the original feature or structural information retained in each view can provide complementary information to the other during subsequent cross-view consistency learning, thereby mitigating this issue. Meanwhile, in addition to high- and low-dimensional contrastive learning, we further introduce clustering-guided learning and jointly optimize it with adaptive denoising in an end-to-end manner. This enables clustering semantics to guide the learning of denoising masks, encouraging the denoising module to adaptively select clustering-relevant information during training, thereby achieving clustering-oriented adaptive denoising rather than merely filtering input information. Furthermore, considering that noise may propagate into the model optimization process, we introduce an importance-aware gradient pruning strategy to suppress potentially unstable parameter updates, further enhancing model robustness from an optimization perspective.

In summary, our ADHERE addresses the propagation of noise throughout hypergraph clustering by achieving multi-level denoising and joint learning across input information, node representations, and model optimization. Meanwhile, we have incorporated references to the relevant works into both Section 2.4 and Section D.2 to further improve the quality of our manuscript.

Q2: Discussions About the Theorems.

R2: Thanks for the comment! To further clarify the connection between our adaptive denoising mechanism and the theoretical analysis, we formulate both the feature and incidence matrices as a noisy matrix M=M⋆+NM=M^\star+N, where M⋆M^\star denotes the underlying clean matrix and NN denotes the noise matrix. According to the adaptive denoising mechanism of ADHERE, the denoised matrix can be expressed as M^=M⊙T\widehat{M}=M\odot T, where TT denotes the learnable denoising mask. Through mathematical derivation, we can obtain that when: ∑i,j(1−Tij)Nij2>∑i,j(1−Tij)(Mij⋆)2,\sum_{i,j}(1-T_{ij})N_{ij}^2>\sum_{i,j}(1-T_{ij})(M_{ij}^\star)^2, we have ∥M^−M⋆∥F<∥M−M⋆∥F.\|\widehat{M}-M^\star\|_F<\|M-M^\star\|_F. This indicates that when the learned mask selectively suppresses noisy information such that the weighted suppression of noise exceeds that of the informative signal, the adaptive denoising mechanism can effectively reduce input perturbations. This is consistent with the motivation behind our joint optimization strategy, which encourages information selection through sparsity regularization while promoting the preservation of clustering-relevant information through cross-view consistency learning and clustering-guided learning. 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 Ratio5%10%20%30%
Feature Perturbation Ratio (After/Before)0.71340.70280.69730.7022
Structural Perturbation Ratio (After/Before)0.79330.80620.86440.8865

As shown in the results, our proposed denoising mechanism effectively suppresses both feature and structural noise across different noise ratios. Combined with Theorems 1 and 2, these results suggest that the denoised views can contribute to improving the quality of node representations and clustering performance.

In Theorem 2, the condition (∣Bs∣<nmin⁡)(|B^s|<n_{\min}) is introduced to ensure a one-to-one correspondence between cluster labels, thereby translating the upper bound on the number of potentially misclustered nodes into an upper bound on the actual clustering error rate. Even if this condition is not satisfied, the upper bound on the number of potentially misclustered nodes in Theorem 2 remains valid, allowing the theoretical analysis to still characterize the impact of representation perturbations on clustering stability.

The two theorems aim to characterize how input perturbations propagate into representation deviations and clustering errors. Their generality gives our theoretical analysis broader significance by providing theoretical guidance for view construction in hypergraph clustering from the perspective of perturbation reduction. Although both theorems are applicable to different two-view hypergraph contrastive learning methods, this does not imply that these methods possess the same ability to suppress perturbations. Guided by this theoretical perspective, our ADHERE constructs complementary adaptive denoising views, establishes the connection between our adaptive denoising mechanism and the perturbation terms in the theoretical analysis, and empirically validates its effectiveness.

Q3: About the Experiment Results.

R3: Thank you for your careful analysis of our experimental results! Our ADHERE achieves the best performance in the majority of the experiments. On several datasets, ADHERE achieves substantial performance improvements. For example, on DBLP, ACC and F1 improve by 6.86 and 7.71 percentage points, while on 20NewsW100, F1 and ARI improve by 7.92 and 7.30 percentage points. These results demonstrate the broad applicability and overall competitiveness of ADHERE across different data scales, feature spaces, and high-order topological structures.

Regarding Cora-A, different methodological designs and inductive biases may lead to distinct advantages on datasets with different characteristics. Notably, as shown in Table 2, removing the feature-denoised view or the structure-denoised view on Cora-A leads to ACC drops of 5.77 and 7.19 percentage points, further validating the effectiveness of our adaptive denoising mechanism on this dataset. Therefore, although different methods may exhibit their respective advantages on specific datasets, both the overall performance comparison and the ablation studies provide strong evidence supporting the effectiveness of our ADHERE.

Q4: About the Ablation Variant "w/o Lden\mathcal{L}_{den}".

R4: Thank you for the comment! In the "w/o Lden\mathcal{L}_{den}" 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. In addition, in Section 4.4, we further validate the effectiveness of the entire adaptive denoising mechanism by comparing the adaptively denoised views with the original views without denoising. Accordingly, we have revised the relevant descriptions in both Section 4.3 and Section E.3 to more accurately clarify the setting and purpose of this ablation study.

We thank you again for your careful review and constructive comments, which help us further improve the quality of our manuscript. All of the revisions mentioned above are highlighted in blue in the revised manuscript, please refer to it for details. We hope our discussions, clarifications and additional experimental results adequately address your concerns.

Reviewer_LCHiOpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The authors propose an adaptive denoising framework for hypergraph clustering that constructs two denoised views using learnable feature and structural masks. Through cross-view contrastive learning, clustering-guided learning and gradient pruning, it improves optimization stability and node representations. Theoretical analyses and experiments demonstrate the effectiveness of the proposed approach.

Critical strengths

The two denoised views provide a reasonable foundation for subsequent representation learning; the learnable masks enable adaptive and end-to-end denoising; the gradient pruning strategy is simple and easy to integrate into existing frameworks; experiments with different hypergraph encoders demonstrate the adaptability of the proposed method across architectures.

Critical weaknesses

The masking-based denoising mechanism cannot directly recover missing hyperedges or correct corrupted features, limiting its ability to handle different types of noise; the pseudo-labels from k-means may be unreliable during early training, potentially reinforcing incorrect assignments and affecting clustering performance; Theorem 2 assumes separable clean node representations, which may not hold for real-world hypergraphs with substantial structural mixing; the gradient pruning strategy lacks a clear explanation of how selectively suppressing parameter updates improves optimization stability.

Questions for authors
  1. How does the proposed method handle different types of noise?
  2. How does the proposed method mitigate the impact of incorrect pseudo-labels?
  3. Can the authors further discuss the assumption of separable clean representations mentioned in theorem 2?
  4. Could the authors explain how selectively suppressing parameter updates improves optimization stability?
1 comment

We greatly appreciate your time and effort in evaluating our manuscript. We address each of your concerns in the following responses.

Q1: How does the proposed method handle different types of noise?

R1: Thank you for your question! Our method handles different types of noise through adaptive information selection and complementary representation learning.

  • For feature noise, the element-wise feature mask suppresses corrupted or clustering-irrelevant features, reducing their influence on node representations. When features are partially missing, hypergraph message passing aggregates available information from other nodes within observed hyperedges, providing contextual signals that can compensate for incomplete node attributes at the representation level.

  • For structural noise, the incidence mask selectively suppresses unreliable node–hyperedge associations while retaining useful connections within the same hyperedge. When associations or hyperedges are missing, the model can still exploit node attributes and remaining high-order relationships to learn cluster membership. Cross-view consistency and clustering-guided learning further encourage the preservation of shared cluster semantics, helping reduce dependence on individual structural connections.

Together, these mechanisms suppress unreliable information and exploit the remaining informative signals to mitigate the impact of missing information on clustering.

Q2: How does the proposed method mitigate the impact of incorrect pseudo-labels?

R2: Thank you for your valuable comment! Our ADHERE mainly mitigates the impact of incorrect pseudo-labels through dynamic pseudo-label updates and joint multi-view optimization. Rather than relying on fixed initial pseudo-labels, our ADHERE dynamically updates the KK-means clustering results as node representations are iteratively optimized, allowing early incorrect assignments to be adjusted during subsequent training. Besides, we construct two complementary views through adaptive denoising to reduce reliance on pseudo-labels from a single view, thereby enhancing the error tolerance. Furthermore, since ADHERE jointly optimizes contrastive learning, clustering-guided learning, and denoising objectives, representation learning does not solely rely on pseudo-label supervision.

In summary, the dynamic pseudo-label updates and joint multi-view optimization mechanism help mitigate the accumulation of incorrect pseudo-labels during iterative training, thereby improving the stability of clustering learning.

Q3: Discussion About the Assumption of Separable Clean Representations.

R3: Thank you for the valuable comment! The separability assumption in Theorem 2 is primarily used to characterize the underlying cluster structure of ideal clean representations, and this assumption is compatible with the class mixing observed in real-world hypergraph representations. Specifically, we assume that the clean node representations have KK cluster centers, with a minimum inter-cluster distance of Δ>0\Delta>0 and within-cluster dispersion characterized by rr. This setting allows a certain degree of within-cluster dispersion and cluster overlap, thereby capturing complex cluster distributions in real-world hypergraphs. When the inter-cluster separation is sufficiently large relative to the within-cluster dispersion and the perturbation of denoised representations is small, Theorem 2 provides a tighter upper bound on the clustering error. For hypergraphs with substantial class mixing, the within-cluster dispersion and representation perturbation may increase accordingly, leading to changes in the theoretical bound that reflect the error characteristics of different cluster structures.

Q4: Could the authors explain how selectively suppressing parameter updates improves optimization stability?

R4: Thank you for your valuable question! After reviewing our implementation, we have revised the description of the parameter selection strategy in Section 2.5 of the manuscript to clarify that parameter importance is evaluated by jointly considering parameter and gradient magnitudes. During training, noise may induce gradient fluctuations, while selectively masking the gradients of less important parameters reduces the number of parameters involved in each update, thereby limiting the impact of potentially unstable updates on model optimization. Therefore, our ADHERE adopts an importance-aware selective gradient update strategy that limits potentially unstable parameter updates while preserving model adaptability, thereby improving optimization stability.

Once again, we sincerely thank you for your valuable comments and constructive suggestions. All of the revisions mentioned above are highlighted in blue in the revised manuscript, please refer to it for details. And we hope our responses can address your concerns.

Reviewer_pt6jOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper studies denoising hypergraph clustering and proposes a method that adopts adaptive denoising, contrastive learning, and clustering guidance. Based on two denoised views constructed from the adaptive denoising strategy, the paper employs cross-view contrastive learning and clustering-guided self-supervised learning. Moreover, the paper also adopts a gradient pruning strategy to improve the optimization stability. Both theoretical analyses and experiments demonstrate the effectiveness of the proposed method.

Critical strengths

--Hypergraphs in real world are often affected by noise, and this paper addresses this from the perspective of feature and structure, making the problem practical. --Compared with random augmentation, this paper employs learnable masks to adaptively select useful feature and structural information, which is reasonable and novel. --This paper provides theoretical analyses on hypergraph propagation stability and clustering error, offering insights into how feature and structural perturbations affect the final clustering.

Critical weaknesses

--In Eq. (1), the feature mask is generated solely from a single node feature value, without incorporating the complete features of a node or its neighborhood information. However, whether a feature is noisy often depends on its context. The same feature value may carry different semantic information across nodes or structural environments, such a masking strategy may limit the model's ability to accurately assess feature reliability. --Although the appendix reports detailed implementation settings, several important hyperparameters, including the continuous relaxation temperature and contrastive learning temperature, are not explicitly specified. --The pseudo-label supervised learning may be sensitive to the quality of the initial clustering results, which may contain incorrect cluster assignments, and this can contaminate the representation in later training epochs. This paper does not discuss how to mitigate this potential issue. --The method uses the absolute magnitude of each parameter as an importance criterion to determine whether a gradient should be updated. However, parameter magnitude does not necessarily reflect gradient reliability, and parameters with small magnitudes may still contribute to important optimization directions.

Questions for authors

Why does Eq. (1) generate feature masks based solely on a single node feature without incorporating the complete node features or its neighborhood information? How the authors consider the risk of progressively reinforcing incorrect pseudo-labels? Does the proposed method include any mechanism to mitigate this issue? Why does the parameter pruning module use parameter magnitudes rather than gradient magnitudes or other gradient statistics to determine which gradients to update, and how reliably can parameter magnitudes reflect gradient quality? Could the authors further clarify the connection between the theoretical analyses and the denoising mechanism, particularly under which conditions this mechanism can effectively reduce these perturbations?

4 comments

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: Discussion About the Feature Mask Generation in Eq. (1).

R1: Thank you for the comment! We adopt element-wise feature masking mainly for the following considerations. First, element-wise masking enables fine-grained selection of individual features. Second, since the original hypergraph may contain both feature and structural noise, directly incorporating neighborhood information into feature mask generation could introduce interference between these two types of noise. Furthermore, although each feature mask is generated from an individual feature value, its learning process is not entirely independent of contextual information. Neighborhood information and clustering semantics can also guide feature mask learning through subsequent modules. Our design aims to achieve feature denoising by combining lightweight, fine-grained feature selection with context-aware representation learning.

We have further clarified the rationale behind this mechanism in Section 2.2 of the revised manuscript. Please refer to it for more details.

Q2: Discussion About the Mechanism to Mitigate the Incorrect Pseudo-label Reinforcement.

R2: Thank you for your valuable question! During training, incorrect pseudo-labels generated from early clustering may affect subsequent representation learning. Our ADHERE mainly mitigates this through the following mechanisms. First, rather than using the initial pseudo-labels as fixed supervision signals, we dynamically update the K-means clustering results as node representations are iteratively optimized, allowing early incorrect assignments to be adjusted during subsequent training. Second, our ADHERE constructs two complementary denoised views to provide complementary optimization constraints for node representations, thereby reducing reliance on pseudo-labels from a single view. Furthermore, our ADHERE jointly optimizes contrastive learning, clustering-guided learning, and denoising objectives, so the representation learning does not solely rely on pseudo-label supervision.

In summary, our ADHERE mitigates the cumulative impact of incorrect pseudo-labels through dynamic pseudo-label updates and joint multi-view optimization.

Q3: Discussion About the Parameter Pruning strategy.

R3: Thank you for your valuable comment! Upon reviewing our actual implementation, we find that our strategy evaluates parameter importance by jointly considering parameter magnitudes and gradient information, i.e., we use the absolute product of each parameter value and its gradient as the importance score to select parameters for gradient updates. This design is motivated by two considerations: firstly, parameter magnitude provides a simple indication of parameter importance, while gradient magnitude reflects the sensitivity of the optimization objective to parameter changes. Their product provides a first-order approximation of the loss variation caused by removing a parameter, thereby offering a more comprehensive importance criterion. This strategy allows the model to focus optimization on parameters with greater estimated influence on the objective, while limiting potentially less informative updates without permanently removing parameters.

We have corrected the parameter selection rule in Section 2.5 of the revised manuscript to accurately reflect our original implementation.

Q4: Clarifications of the Connection Between the Theoretical Analyses and the Denoising Mechanism.

R4: Thank you for your valuable comment! To further clarify the connection between our adaptive denoising mechanism and the theoretical analysis, we formulate both the feature and incidence matrices as a noisy matrix M=M⋆+NM=M^\star+N, where M⋆M^\star denotes the underlying clean matrix and NN denotes the noise matrix. According to the adaptive denoising mechanism of ADHERE, the denoised matrix can be expressed as M^=M⊙T\widehat{M}=M\odot T, where TT denotes the learnable denoising mask. Through mathematical derivation, we can obtain that when: ∑i,j(1−Tij)Nij2>∑i,j(1−Tij)(Mij⋆)2,\sum_{i,j}(1-T_{ij})N_{ij}^2>\sum_{i,j}(1-T_{ij})(M_{ij}^\star)^2, we have ∥M^−M⋆∥F<∥M−M⋆∥F.\|\widehat{M}-M^\star\|_F<\|M-M^\star\|_F. This indicates that when the learned mask selectively suppresses noisy information such that the weighted suppression of noise exceeds that of the informative signal, the adaptive denoising mechanism can effectively reduce input perturbations. To further verify whether the learned denoising masks of 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 Ratio5%10%20%30%
Feature Perturbation Ratio (After/Before)0.71340.70280.69730.7022
Structural Perturbation Ratio (After/Before)0.79330.80620.86440.8865

As shown in the results, our proposed denoising mechanism effectively suppresses both feature and structural noise across different noise ratios. Combined with Theorems 1 and 2, these results suggest that the denoised views can contribute to improving the quality of node representations and clustering performance.

Q5: About the Implementation Details.

R5: Thank you for your valuable comment! We have further supplemented the implementation details in Appendix E.2 of the revised manuscript, including the specific settings of the continuous relaxation temperature τd\tau_d and contrastive learning temperature τh\tau_h, along with clarification of their usage during training. We have also reviewed the remaining hyperparameter configurations to ensure the reproducibility of our method.

We sincerely appreciate your valuable feedback and thoughtful suggestions. All of the revisions mentioned above are highlighted in blue in the revised manuscript, please refer to it for details. We hope our detailed responses and revisions have adequately addressed your concerns.

Reviewer_pt6jOpenReview ↗

Concerns have been addressed

All my concerns have been addressed, and I am happy to recommend acceptance.

Thanks for your valuable feedback and recognition of our work, and we're happy that our responses have addressed your concerns.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.