acceptodds
Under review as a conference paper at ICLR 2027

HyperRoGMAE: Multi-Scale Semantic Hypergraphs for Robust Graph Masked Autoencoder against Structural Attacks

Abstract

Graph Masked AutoEncoders (GMAEs) propagate contextual information through the observed pairwise topology to reconstruct masked graph elements. This paradigm makes them vulnerable to structural attacks that corrupt the observed topology. Existing robust GMAEs mitigate this vulnerability through adversarial masking, structural augmentation, and edge-level filtering or reweighting. However, their contextual modeling relies primarily on pairwise relations and thus lacks multi-scale, higher-order semantic context for robust reconstruction under structural attacks. To this end, we propose , a graph-guided bust raph asked utoncoder that constructs a multi-scale semantic hypergraph for masked representation learning. Specifically, HyperRoGMAE first constructs a multi-scale semantic hypergraph and estimates semantic reliability based on assignment margins and cross-view stability. It then pretrains a hypergraph teacher that encodes masked views and provides reliability-weighted higher-order guidance. During student pretraining, it attenuates suspicious edges according to feature similarity and hypergraph-membership consistency. Finally, it introduces a node-aware adaptive fusion strategy to fuse graph and semantic representations and employs dual reconstruction objectives to preserve their complementary information. Extensive experiments on five datasets demonstrate the effectiveness of the proposed HyperRoGMAE under four structural attacks.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph Masked Autoencoders, Semantic Hypergraphs, Adversarial Attacks, Robustness

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_Ana8OpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes HyperRoGMAE, a robust graph masked autoencoder designed to mitigate structural attacks through multi-scale semantic hypergraphs. It constructs overlapping attribute-derived hyperedges using clustering at multiple resolutions and estimates node reliability through assignment margins and cross-view stability. A pretrained hypergraph teacher provides contextual semantic representations and reliability-weighted hyperedge prompts. The graph student attenuates potentially corrupted edges using feature similarity and hypergraph-membership consistency. Node-aware adaptive fusion integrates graph and semantic representations during pretraining, while dual reconstruction objectives supervise representation learning. Experiments on several graph datasets evaluate node classification under various experimental settings and tasks validate the superiority of the proposed model.

Critical strengths
  1. The paper identifies a meaningful limitation of topology-dependent masked graph learning. When structural attacks corrupt the observed neighborhoods, reconstructing masked features from those neighborhoods can become unreliable. Thus, the paper is well-motivated, whose motivation is clearly presented in Fig. 1.

  2. The experimental analysis is rich. The empirical evaluation considers both global and targeted structural attacks, with separate poisoning and evasion settings. The reported results show relatively stable performance under stronger perturbations. The appendix also includes more detailed component comparisons and reliability diagnostics beyond the main accuracy tables.

Critical weaknesses
  1. A major claimed contribution is node-aware adaptive fusion between graph and semantic representations. However, Eq. 2121 shows that downstream evaluation uses only the graph-student representation, excluding the fused representation, semantic teacher, and fusion gate. Thus, adaptive fusion acts as a training-time supervision mechanism rather than a component of the final predictor. I would encourage the authors to have more explanations on this point.

  2. According to Eqs. 16–18, greater teacher–student disagreement increases the semantic fusion weight when other inputs are fixed. However, I think that disagreement alone cannot establish whether the graph student or the semantic teacher is more reliable. If the teacher is inaccurate, increasing semantic guidance may propagate erroneous information.

  3. The semantic hypergraph, node reliability, and edge reliability all depend on the original node attributes. This makes the framework robust to topology changes only insofar as the attribute-derived reference remains trustworthy. As the authors acknowledge in Appendix A.6, feature corruption could directly undermine the semantic reference, while heterophilic graphs could cause legitimate cross-class edges to receive low weights.

Questions for authors
  1. Why does downstream inference use only the graph student rather than the adaptively fused representation?

  2. What prevents the fusion gate from assigning larger weights to an inaccurate semantic teacher when teacher–student disagreement increases?

  3. Could the authors discuss the potential methods to resolve the problem of attribute-based reliability?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Ana8 (W1-W2)

Thank you for your time and for recognizing the strength of this paper. We have marked the revisions on downstream representations and adaptive fusion in blue in the revised manuscript. The shared experiments on feature deletion and heterophilous graphs are reported in Tables 3 and 9, with their captions and accompanying discussion marked in magenta.

W1. A major claimed contribution is node-aware adaptive fusion between graph and semantic representations. However, Eq.21 shows that downstream evaluation uses only the graph-student representation, excluding the fused representation, semantic teacher, and fusion gate. Thus, adaptive fusion acts as a training-time supervision mechanism rather than a component of the final predictor. I would encourage the authors to have more explanations on this point.
Q1. Why does downstream inference use only the graph student rather than the adaptively fused representation?

Thank you for your careful comments. The adaptive fusion module is used during Stage II pretraining, whereas the reported downstream linear probe uses the graph-student representation ZG\mathbf Z_G. Specifically, the fused reconstruction loss propagates gradients through ZF\mathbf Z_F to train the graph student with semantic guidance. The separate graph reconstruction loss directly trains ZG\mathbf Z_G, allowing the student to retain information from the reliability-weighted observed graph. Thus, the teacher and fusion gate guide training but are not part of the embedding used by the reported downstream classifier. The attribute-derived semantic hypergraph still affects ZG\mathbf Z_G through the observed-edge reliability weights.

To examine the choice of downstream representation, we additionally evaluated graph-student, fused, and semantic-teacher embeddings from the trained models under the same data splits and linear-probe protocol. The graph-student representation has the highest mean accuracy in all six settings, providing empirical support for its use in the reported downstream protocol.

SettingDatasetGraph studentFusedSemantic teacher
25% Metattack poisoningCora78.59 ± 0.7270.65 ± 5.8164.59 ± 0.52
Citeseer72.76 ±0.5371.89 ± 0.7868.01 ± 0.54
PubMed84.09 ± 0.0480.70 ± 0.6579.25 ± 0.84
CleanCora83.48 ± 0.4672.91 ± 7.8064.64 ± 0.51
Citeseer75.50 ± 0.1272.66 ± 0.6468.01 ± 0.54
PubMed84.69 ± 0.0380.46 ± 0.8979.25 ± 0.84

W2. According to Eqs. 16-18, greater teacher-student disagreement increases the semantic fusion weight when other inputs are fixed. However, I think that disagreement alone cannot establish whether the graph student or the semantic teacher is more reliable. If the teacher is inaccurate, increasing semantic guidance may propagate erroneous information.
Q2. What prevents the fusion gate from assigning larger weights to an inaccurate semantic teacher when teacher-student disagreement increases?

Thank you for your insightful observation. We agree that teacher-student disagreement alone cannot determine which branch is more reliable for a given node. With the other gate inputs fixed, greater disagreement increases the semantic fusion weight, even when the teacher is inaccurate. Our motivation comes from the topology-only threat model that node attributes remain unchanged, and the teacher does not use the perturbed adjacency. The teacher thus provides a complementary semantic reference when the observed topology is corrupted.

The gate also incorporates attribute-derived semantic reliability, and the separate graph reconstruction loss gives the student a direct training signal independent of the fused representation. These design choices reduce the student’s reliance on semantic guidance during pretraining, but the gate can still assign a high weight to an inaccurate teacher. Downstream classification uses only the graph-student embedding ZG\mathbf{Z}_G. The reported results show its robustness under the evaluated structural attacks, while erroneous teacher guidance can still affect how ZG\mathbf{Z}_G is learned.

AuthorsOpenReview ↗

Response to Reviewer Ana8 (W3)

W3. The semantic hypergraph, node reliability, and edge reliability all depend on the original node attributes. This makes the framework robust to topology changes only insofar as the attribute-derived reference remains trustworthy. As the authors acknowledge in Appendix A.6, feature corruption could directly undermine the semantic reference, while heterophilic graphs could cause legitimate cross-class edges to receive low weights.
Q3. Could the authors discuss the potential methods to resolve the problem of attribute-based reliability?

Thank you for your valuable suggestions. We agree that the quality of node attributes affects the semantic hypergraph, node reliability, and edge weights. We evaluated this dependence under random feature deletion and examined performance on heterophilous graphs. The results below are the same as those reported in our earlier responses.

(1) Random feature deletion. We randomly delete 10%–50% of features under 25% Metattack poisoning. The corrupted features are used throughout semantic hypergraph construction, edge weighting, pretraining, and evaluation. HyperRoGMAE achieves the highest mean accuracy in 12 of 15 settings, including all tested deletion rates on Cora and Citeseer. On Photo, DGMAE performs better at 30%-50% deletion. These results show that HyperRoGMAE remains competitive under the tested joint corruption, with reduced accuracy and greater variability on Photo dataset at high deletion rates.

DatasetMethod10%20%30%40%50%
CoraDGMAE’2555.00 ± 0.5353.83 ± 0.6352.50 ± 1.0050.87 ± 0.9849.65 ± 0.35
ArmorGAE’2637.69 ± 13.1347.17 ± 11.0344.92 ± 13.6050.15 ± 0.8550.91 ± 0.15
HyperRoGMAE76.04 ± 0.9873.90 ± 1.0370.56 ± 1.0967.98 ± 1.1462.47 ± 1.44
CiteseerDGMAE’2563.93 ± 1.1763.60 ± 0.5662.66 ± 0.2862.73 ± 0.1162.03 ± 0.29
ArmorGAE’2657.10 ± 15.9556.33 ± 16.9265.34 ± 0.7663.23 ± 1.4263.86 ± 0.38
HyperRoGMAE72.98 ± 0.3771.86 ± 0.4470.41 ± 0.9768.62 ± 0.7665.48 ± 0.76
PhotoDGMAE’2590.36 ± 0.5190.06 ± 0.3590.00 ± 0.3289.62 ± 0.4489.42 ± 0.55
ArmorGAE’2639.15 ± 4.2535.83 ± 5.5735.43 ± 3.5936.28 ± 2.6734.48 ± 2.53
HyperRoGMAE91.52 ± 0.3490.92 ± 0.6489.67 ± 1.2387.12 ± 5.3883.72 ± 6.32

(2) Heterophilous graphs. HyperRoGMAE achieves the best or tied-best accuracy in 13 of 16 settings. Under 25% Metattack, it outperforms the strongest compared baseline on Actor, Cornell, and Wisconsin datasets, while ArmorGAE performs better on Texas. These results demonstrate competitive performance on the evaluated heterophilous graphs.

DatasetMethodCleanMeta 5%Meta 15%Meta 25%
TexasMaskGAE’2364.8648.6564.8662.16
AUGMAE’2451.3559.4651.3554.05
ArmorGAE’2659.4662.1662.1667.57
HyperRoGMAE62.1659.4664.8662.16
ActorMaskGAE’2329.0126.2527.2426.51
AUGMAE’2425.1325.2025.4625.13
ArmorGAE’2628.6226.0526.1825.66
HyperRoGMAE30.3329.8030.4630.20
CornellMaskGAE’2337.8437.8437.8437.84
AUGMAE’2448.6559.4637.8437.84
ArmorGAE’2645.9545.9537.8437.84
HyperRoGMAE54.0559.4651.3551.35
WisconsinMaskGAE’2347.0647.0652.9445.10
AUGMAE’2458.8254.9047.0645.10
ArmorGAE’2650.9856.8652.9449.00
HyperRoGMAE66.6768.6364.7154.90

(3) Potential improvements. For feature corruption, we plan to construct the semantic hypergraph from memberships that recur across multiple denoised feature views and reduce the contribution of uncertain memberships to teacher guidance and edge weighting. For heterophilous graphs, we plan to learn edge compatibility from recurring neighborhood patterns so that legitimate edges between dissimilar nodes can retain weight when supported by their structural context.

Thank you for the careful response and addtional experimental results that solved most of my problems. I will keep my positive rating and tend to accept the paper.

Thank you for your follow-up and positive assessment. We are glad that our responses and additional experiments addressed most of your concerns. We sincerely appreciate your positive feedback.

Reviewer_yRP4OpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes HyperRoGMAE, a robust graph masked autoencoder against structural attacks. It constructs multi-scale semantic hypergraphs from node attributes and pretrains a hypergraph teacher to provide topology-independent guidance. The graph student incorporates reliability-based edge reweighting, node-aware adaptive fusion, and dual reconstruction objectives. Experiments on five datasets under four structural attacks demonstrate improved robustness in most settings.

Critical strengths
  • The paper is well-motivated and easy to follow.
  • HyperRoGMAE achieves competitive or superior performance across five datasets and four structural attacks, with particularly substantial improvements under severe perturbations.
Critical weaknesses
  • The main components, including multi-scale K-means clustering, similarity-based edge reweighting, teacher-student guidance, and adaptive fusion, largely build on established techniques. The novelty appears marginal and primarily in their integration.
  • The paper does not adequately demonstrate that multi-scale hypergraphs provide advantages over simpler semantic structures, such as feature-based kNN graphs or single-scale clustering. This weakens the justification for the proposed higher-order modeling.
  • The ablation study shows that edge reliability weighting contributes the largest performance gains. However, comparisons against a standalone edge-reweighted GraphMAE and simpler semantic guidance mechanisms are missing, making it difficult to isolate the benefits of the hypergraph teacher and adaptive fusion.
Questions for authors

Please refer th weaknesses above.

5 comments
AuthorsOpenReview ↗

Response to Reviewer yRP4 (W1)

We appreciate your time and the recognition of the strength of our work. We have marked the revision in the color orange on the paper.

W1. The main components, including multi-scale K-means clustering, similarity-based edge reweighting, teacher-student guidance, and adaptive fusion, largely build on established techniques. The novelty appears marginal and primarily in their integration.

Thank you for your valuable suggestion. The key contribution of HyperRoGMAE is a topology-independent semantic reference that coordinates context selection and teacher guidance for robust masked reconstruction. This design addresses unreliable structural context and accounts for uncertainty in attribute-derived semantic groups. This motivation was also recognized by Reviewer 8zZo as “a valid research direction” and Reviewer Ana8 as addressing “a meaningful limitation of topology-dependent masked graph learning.” Specifically:

  1. Coordinated context selection and propagation. The same overlapping hypergraph memberships guide node masking and contribute to edge weights. Reconstruction-context selection and message propagation therefore share a common semantic reference.
  2. Confidence-weighted semantic transfer. Assignment margins and cross-view stability estimate semantic confidence. This confidence weights the teacher’s contextual representations, allowing nodes to contribute differently to semantic guidance.
  3. Node-dependent reconstruction guidance. Fusion incorporates semantic confidence and teacher-student disagreement. The fused and graph-only reconstruction objectives jointly train the graph student, whose representation is used downstream.

As summarized in the comparison table, ProtoMGAE emphasizes prototype-based clustering consistency. ACGMAE contrasts a feature-based kNN graph with the attacked graph and ArmorGAE combines semantic topology augmentation with adversarial edge reconstruction. HyperRoGMAE uses the same semantic organization to guide multiple decisions within reconstruction.

MethodAuxiliary referenceHow it guides learning
ProtoMGAE’24Learned prototypes and an online-target networkView alignment and prototype-aware clustering consistency
ACGMAE’25Feature-based kNN graphAdversarial feature masking and cross-graph contrast with estimated positive-pair probabilities
ArmorGAE’26Graph augmentation based on feature similarity and degree constraintsTopology augmentation and adversarial edge reconstruction
HyperRoGMAEAttribute-derived, multi-scale overlapping hypergraphShared memberships guide masking and edge weighting. Semantic confidence calibrates teacher context and node-wise fusion

The new comparisons valid whether simpler alternatives can achieve similar performance. In Table 14, the multi-scale hypergraph achieves higher accuracy than each single-scale variant in all 12 evaluated settings and outperforms the kNN prior in 10 of 12 settings. In Table 13, on Cora, HyperRoGMAE’s reported mean accuracy exceeds that of the kNN-guidance and fixed-fusion variants by 2.36%-3.13% and 0.51%-1.25%, respectively, across the three Metattack poisoning rates.

AuthorsOpenReview ↗

Response to Reviewer yRP4 (W2-W3)

W2. The paper does not adequately demonstrate that multi-scale hypergraphs provide advantages over simpler semantic structures, such as feature-based kNN graphs or single-scale clustering. This weakens the justification for the proposed higher-order modeling.

Thank you for your valuable suggestion. We have added comparisons with a feature-based kNN graph and single-scale hypergraphs constructed using 16, 32, 64, or 128 clusters. We evaluate these semantic structures under both poisoning and evasion attacks at 5%, 15%, and 25% perturbation rates. As shown in the table, HyperRoGMAE achieves higher accuracy than each tested single-scale hypergraph in all 12 evaluated settings, and outperforms the feature-based kNN graph in 10 of them. These results suggest that combining clusters at different resolutions can provide more useful semantic guidance than relying on any one of the tested resolutions. The accuracy achieved with the feature-based kNN graph is slightly higher than that of HyperRoGMAE on PubMed under 15% and 25% poisoning. A possible explanation is that local feature similarity already provides effective semantic guidance on PubMed, leaving limited additional benefit from combining multiple cluster resolutions in these settings.

DatasetAttack settingkNN graphSingle-scale 16Single-scale 32Single-scale 64Single-scale 128HyperRoGMAE (multi-scale)
CoraPoisoning (5% / 15% / 25%)77.70 / 76.49 / 75.4679.51 / 77.59 / 74.9678.61 / 76.71 / 76.0279.47 / 77.95 / 76.9279.68 / 78.30 / 76.8980.46 / 78.85 / 78.59
Evasion (5% / 15% / 25%)78.48 / 77.98 / 77.4379.94 / 78.68 / 78.8179.73 / 78.97 / 77.9879.49 / 79.49 / 78.8579.19 / 79.74 / 79.1080.06 / 79.95 / 79.77
PubMedPoisoning (5% / 15% / 25%)84.50 / 84.47 / 84.3181.64 / 81.24 / 81.0082.07 / 81.83 / 81.6682.25 / 82.04 / 81.5783.44 / 83.12 / 82.9384.57 / 84.27 / 84.09
Evasion (5% / 15% / 25%)84.42 / 84.15 / 83.9881.60 / 81.33 / 81.0582.26 / 81.97 / 81.9182.34 / 81.93 / 81.7683.21 / 83.13 / 83.0484.61 / 84.55 / 84.34

W3. The ablation study shows that edge reliability weighting contributes the largest performance gains. However, comparisons against a standalone edge-reweighted GraphMAE and simpler semantic guidance mechanisms are missing, making it difficult to isolate the benefits of the hypergraph teacher and adaptive fusion.

Thank you for your valuable suggestions. We expanded the ablation study with a standalone edge-reweighted GraphMAE baseline, a kNN guidance variant with adaptive fusion, and a hypergraph teacher variant with fixed fusion. Results are reported for 5%, 15%, and 25% Metattack poisoning, in that order. The values in parentheses give HyperRoGMAE’s accuracy minus that of each variant. Specifically, the reported HyperRoGMAE accuracy on Cora exceed those of the kNN guidance and fixed-fusion controls by 2.36%-3.13% and 0.51%-1.25%, respectively. These comparisons suggest that the hypergraph teacher and adaptive fusion can each yield additional gains. The fixed fusion variant outperforms HyperRoGMAE by 1.00% in terms of accuracy under 25% Metattack on Citeseer. Furthermore, the kNN guidance baseline performs within a 0.08% margin of HyperRoGMAE’s reported results on PubMed, suggesting limited additional benefit from the hypergraph teacher in this comparison.

DatasetGraphMAE + Edge WeightskNN Guidance + Adaptive FusionHypergraph Teacher + Fixed FusionOurs
Cora75.19 (+5.27) / 72.82 (+6.03) / 68.57 (+10.02)77.70 (+2.76) / 76.49 (+2.36) / 75.46 (+3.13)79.95 (+0.51) / 77.65 (+1.20) / 77.34 (+1.25)80.46 / 78.85 / 78.59
Citeseer74.88 (+0.08) / 73.88 (+0.31) / 73.00 (−0.24)72.18 (+2.78) / 71.38 (+2.81) / 69.78 (+2.98)74.39 (+0.57) / 73.94 (+0.25) / 73.76 (−1.00)74.96 / 74.19 / 72.76
PubMed86.18 (−1.61) / 86.11 (−1.84) / 85.90 (−1.81)84.50 (+0.07) / 84.19 (+0.08) / 84.01 (+0.08)83.78 (+0.79) / 83.59 (+0.68) / 83.75 (+0.34)84.57 / 84.27 / 84.09
AuthorsOpenReview ↗

Follow-up on Our Response and Additional Comparisons

Dear Reviewer yRP4,

Thank you again for your constructive suggestions. In our response posted on November 12, we clarified the contribution of our coordinated semantic guidance and added comparisons with kNN and single-scale priors, standalone edge-reweighted GraphMAE, and simpler guidance and fusion variants. We also discussed the settings where simpler alternatives remain competitive.

As the discussion period is closing soon, could you please let us know whether these comparisons address your concerns about the justification and contribution of our design? If so, would you be willing to reconsider your assessment and rating?

We would appreciate any update and would be glad to discuss any remaining questions. Thank you for your time.

Reviewer_8zZoOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how to enable a Graph Masked AutoEncoder to learn discriminative node representations even when the graph structure is subjected to adversarial attacks. The authors propose HyperRoGMAE, whose core idea is to construct a multiscale semantic hypergraph using node attributes, in addition to relying on those edges. This hypergraph aids in node feature reconstruction and helps identify suspicious edges. The model achieved significant performance improvements under various structural attacks, demonstrating particularly strong stability compared to standard GraphMAE and some Robust GMAE variants, especially under severe attacks. However, the experiments designed in this paper had a relatively limited scope of attacks and dataset coverage, and thus do not represent a broad range of graph learning scenarios.

Critical strengths
  1. Graph-structure attacks can corrupt the GMAE’s reconstruction context. Establishing auxiliary references based on node attributes that remain invariant under structural attacks is a valid research direction.
  2. Significant gains were achieved under strong Metattack, Nettack, Random, and DICE settings across multiple datasets.
  3. Diagnostics for reliability mechanisms are provided.
Critical weaknesses
  1. According to the method, HyperRoGMAE is built using complete, unmasked node attributes. Did the compared baselines also use this setup? Is the evaluation comparison fair?
  2. Could the authors discuss whether the model remains competitive when attack methods modify node attributes?
  3. The authors only conducted validation on homogeneous graphs. How does the model perform on heterophilous graphs, like Texas or Actor.
  4. HyperRoGMAE includes SVD, K-means, teacher training, and student training, but the authors report only the average training time for the student, making it difficult to assess the time consumption of the entire framework.
Questions for authors

Please refer to the Weaknesses.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 8zZo (W1-W2)

We appreciate your detailed review and the recognition of the strength of our work. We have marked the revision in the color magenta on the paper.

W1. According to the method, HyperRoGMAE is built using complete, unmasked node attributes. Did the compared baselines also use this setup? Is the evaluation comparison fair?

Thank you for your commons. In this paper, all baselines use the same original node features XX, graphs, and data splits. Structure-oriented baselines, such as MaskGAE (Li et al., 2023) and Bandana (Zhao et al., 2024), use original features to reconstruct structural information. Feature-reconstruction baselines, such as GraphMAE (Hou et al., 2022), AUGMAE (Wang et al., 2024a), and ProtoMGAE (Zheng & Jia, 2024), mask features and reconstruct the originals. HyperRoGMAE constructs the semantic hypergraph from the original node features XX before masking and also uses XX to compute edge weights.

To test whether the results depend on using complete features to build the semantic prior, we split the feature dimensions into two halves. Both variants reconstruct the same half. The full-feature variant uses both halves to construct the semantic hypergraph and compute edge weights, whereas the isolated-view variant uses only the other half. These findings suggest that the isolated-view variant retains mean accuracy on Cora and Citeseer under both clean and poisoned settings. On Photo, using both feature halves to construct the hypergraph and compute edge weights improves mean accuracy by 0.51%-0.77%.

DatasetGraphFull-XX priorIsolated-view priorΔ\Delta
CoraClean76.20 ± 1.0677.14 ± 0.78+0.95 ± 0.95
Metattack 25%73.66 ± 0.9573.96 ± 0.72+0.30 ± 0.24
CiteseerClean69.39 ± 0.6370.15 ± 0.85+0.76 ± 0.57
Metattack 25%67.68 ± 1.6068.63 ± 1.15+0.95 ± 2.10
PhotoClean93.57 ± 0.0493.05 ± 0.26-0.51 ± 0.29
Metattack 25%92.38 ± 0.1791.61 ± 0.33-0.77 ± 0.49

W2. Could the authors discuss whether the model remains competitive when attack methods modify node attributes?

Thank you for your commons. We compared the classification accuracy of HyperRoGMAE with that of DGMAE and ArmorGAE under 25% Metattack poisoning and random deletion of 10%-50% of features. We utilize the corrupted features throughout semantic-hypergraph construction, edge weighting, pretraining, and evaluation. HyperRoGMAE achieves the highest mean accuracy in 12 of 15 settings, including all five deletion ratios on Cora and Citeseer datasets. On Photo dataset, DGMAE performs better at 30%-50% deletion. These results show that HyperRoGMAE remains competitive under the tested joint corruption, with reduced accuracy and greater variability on Photo dataset at high deletion rates.

DatasetMethod10%20%30%40%50%
CoraDGMAE’2555.00±0.5353.83±0.6352.50±1.0050.87±0.9849.65±0.35
ArmorGAE’2637.69±13.1347.17±11.0344.92±13.6050.15±0.8550.91±0.15
HyperRoGMAE76.04±0.9873.90±1.0370.56±1.0967.98±1.1462.47±1.44
CiteseerDGMAE’2563.93±1.1763.60±0.5662.66±0.2862.73±0.1162.03±0.29
ArmorGAE’2657.10±15.9556.33±16.9265.34±0.7663.23±1.4263.86±0.38
HyperRoGMAE72.98±0.3771.86±0.4470.41±0.9768.62±0.7665.48±0.76
PhotoDGMAE’2590.36±0.5190.06±0.3590.00±0.3289.62±0.4489.42±0.55
ArmorGAE’2639.15±4.2535.83±5.5735.43±3.5936.28±2.6734.48±2.53
HyperRoGMAE91.52±0.3490.92±0.6489.67±1.2387.12±5.3883.72±6.32
AuthorsOpenReview ↗

Response to Reviewer 8zZo (W3-W4)

W3. The authors only conducted validation on homogeneous graphs. How does the model perform on heterophilous graphs, like Texas or Actor.

Thank you for your commons. We added evaluations on Texas, Actor, Cornell, and Wisconsin datasets under clean graph and 5%, 15%, and 25% Metattack poisoning, comparing with MaskGAE, AUGMAE, and ArmorGAE. HyperRoGMAE achieves the highest accuracy in 13 of 16 settings. Under 25% Metattack, it exceeds the strongest compared baseline by 3.69%, 13.51%, and 5.90% on Actor, Cornell, and Wisconsin datasets, respectively. On Texas dataset, it ties for the best result at 15% poisoning but trails ArmorGAE at 25%. These results demonstrate competitive performance on the evaluated heterophilous graphs.

DatasetMethodCleanMeta 5%Meta 15%Meta 25%
TexasMaskGAE’2364.8648.6564.8662.16
AUGMAE’2451.3559.4651.3554.05
ArmorGAE’2659.4662.1662.1667.57
HyperRoGMAE62.1659.4664.8662.16
ActorMaskGAE’2329.0126.2527.2426.51
AUGMAE’2425.1325.2025.4625.13
ArmorGAE’2628.6226.0526.1825.66
HyperRoGMAE30.3329.8030.4630.20
CornellMaskGAE’2337.8437.8437.8437.84
AUGMAE’2448.6559.4637.8437.84
ArmorGAE’2645.9545.9537.8437.84
HyperRoGMAE54.0559.4651.3551.35
WisconsinMaskGAE’2347.0647.0652.9445.10
AUGMAE’2458.8254.9047.0645.10
ArmorGAE’2650.9856.8652.9449.00
HyperRoGMAE66.6768.6364.7154.90

W4. HyperRoGMAE includes SVD, K-means, teacher training, and student training, but the authors report only the average training time for the student, making it difficult to assess the time consumption of the entire framework.

Thank you for your commons. Appendix A.3.2 describes the complexity of each stage. Using the manuscript notation and treating solver iterations and the number of scales as fixed, truncated SVD, multi-scale mini-batch kk-means, and assignment stability evaluation cost O(NDd′)\mathcal{O}(NDd'), O(Nd′M)\mathcal{O}(Nd'M), and O(TNDd′+TNd′M)\mathcal{O}(TNDd' + TNd'M), respectively. These operations are part of the one-time semantic-prior construction. Computing edge reliability weights with sparse memberships costs O(ED+EκS)\mathcal{O}(ED + E\kappa S). Let B=RH(D+d)+Nd(D+d)B = R_H(D+d) + Nd(D+d). With fixed encoder depths and cached edge weights, one teacher epoch costs O(B)\mathcal{O}(B), and one student epoch costs O(B+Ed)\mathcal{O}(B + Ed), including the frozen teacher and contextual prompting. With fixed epoch budgets, the two training stages together cost O(RH(D+d)+Nd(D+d)+Ed)\mathcal{O}(R_H(D+d) + Nd(D+d) + Ed), with semantic-prior construction accounted for separately.

Taking Cora as an example, semantic-prior construction takes 10.06 s, teacher training takes 2.83 s for 200 epochs (14.2 ms/epoch), and student training takes 6.31 s for 300 epochs (21.0 ms/epoch). Total pretraining takes 19.23 s, including model initialization.

AuthorsOpenReview ↗

Follow-up on Our Response and Additional Experiments

Dear Reviewer 8zZo,

Thank you again for your constructive feedback. In our response posted on November 12, we clarified how node features are used by HyperRoGMAE and the compared baselines, and added an isolated-view analysis to examine the effect of using complete features to construct the semantic prior. We also added evaluations under feature deletion and on heterophilous graphs, together with a breakdown of the full pretraining time.

As the discussion period is closing soon, could you please let us know whether any concerns remain after considering these clarifications and additional results? If they address your concerns, would you be willing to reconsider your assessment and rating?

We would appreciate any update and would be happy to provide further clarification before the discussion closes. Thank you for your time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.