- 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
-
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.
-
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
-
A major claimed contribution is node-aware adaptive fusion between graph and semantic representations. However, Eq. 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.
-
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.
-
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
-
Why does downstream inference use only the graph student rather than the adaptively fused representation?
-
What prevents the fusion gate from assigning larger weights to an inaccurate semantic teacher when teacher–student disagreement increases?
-
Could the authors discuss the potential methods to resolve the problem of attribute-based reliability?
4 comments
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.
| Dataset | Method | 10% | 20% | 30% | 40% | 50% |
|---|---|---|---|---|---|---|
| Cora | DGMAE’25 | 55.00 ± 0.53 | 53.83 ± 0.63 | 52.50 ± 1.00 | 50.87 ± 0.98 | 49.65 ± 0.35 |
| ArmorGAE’26 | 37.69 ± 13.13 | 47.17 ± 11.03 | 44.92 ± 13.60 | 50.15 ± 0.85 | 50.91 ± 0.15 | |
| HyperRoGMAE | 76.04 ± 0.98 | 73.90 ± 1.03 | 70.56 ± 1.09 | 67.98 ± 1.14 | 62.47 ± 1.44 | |
| Citeseer | DGMAE’25 | 63.93 ± 1.17 | 63.60 ± 0.56 | 62.66 ± 0.28 | 62.73 ± 0.11 | 62.03 ± 0.29 |
| ArmorGAE’26 | 57.10 ± 15.95 | 56.33 ± 16.92 | 65.34 ± 0.76 | 63.23 ± 1.42 | 63.86 ± 0.38 | |
| HyperRoGMAE | 72.98 ± 0.37 | 71.86 ± 0.44 | 70.41 ± 0.97 | 68.62 ± 0.76 | 65.48 ± 0.76 | |
| Photo | DGMAE’25 | 90.36 ± 0.51 | 90.06 ± 0.35 | 90.00 ± 0.32 | 89.62 ± 0.44 | 89.42 ± 0.55 |
| ArmorGAE’26 | 39.15 ± 4.25 | 35.83 ± 5.57 | 35.43 ± 3.59 | 36.28 ± 2.67 | 34.48 ± 2.53 | |
| HyperRoGMAE | 91.52 ± 0.34 | 90.92 ± 0.64 | 89.67 ± 1.23 | 87.12 ± 5.38 | 83.72 ± 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.
| Dataset | Method | Clean | Meta 5% | Meta 15% | Meta 25% |
|---|---|---|---|---|---|
| Texas | MaskGAE’23 | 64.86 | 48.65 | 64.86 | 62.16 |
| AUGMAE’24 | 51.35 | 59.46 | 51.35 | 54.05 | |
| ArmorGAE’26 | 59.46 | 62.16 | 62.16 | 67.57 | |
| HyperRoGMAE | 62.16 | 59.46 | 64.86 | 62.16 | |
| Actor | MaskGAE’23 | 29.01 | 26.25 | 27.24 | 26.51 |
| AUGMAE’24 | 25.13 | 25.20 | 25.46 | 25.13 | |
| ArmorGAE’26 | 28.62 | 26.05 | 26.18 | 25.66 | |
| HyperRoGMAE | 30.33 | 29.80 | 30.46 | 30.20 | |
| Cornell | MaskGAE’23 | 37.84 | 37.84 | 37.84 | 37.84 |
| AUGMAE’24 | 48.65 | 59.46 | 37.84 | 37.84 | |
| ArmorGAE’26 | 45.95 | 45.95 | 37.84 | 37.84 | |
| HyperRoGMAE | 54.05 | 59.46 | 51.35 | 51.35 | |
| Wisconsin | MaskGAE’23 | 47.06 | 47.06 | 52.94 | 45.10 |
| AUGMAE’24 | 58.82 | 54.90 | 47.06 | 45.10 | |
| ArmorGAE’26 | 50.98 | 56.86 | 52.94 | 49.00 | |
| HyperRoGMAE | 66.67 | 68.63 | 64.71 | 54.90 |
(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.
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.
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. Specifically, the fused reconstruction loss propagates gradients through ZF to train the graph student with semantic guidance. The separate graph reconstruction loss directly trains ZG, 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 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.
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. The reported results show its robustness under the evaluated structural attacks, while erroneous teacher guidance can still affect how ZG is learned.