acceptodds
Under review as a conference paper at ICLR 2027

DeFence: Robust Node Classification under Joint Label-Structure Noise

Abstract

How can we accurately classify nodes in graphs when both labels and edges are corrupted? Node classification under joint label and structure corruption is important in applications such as fraud detection and citation analysis. For example, in citation networks, errors in automated paper tagging may assign incorrect labels to papers, while errors in reference extraction may introduce spurious links between them. However, most prior works tackle only one noise source, over-trusting structure for label noise and labels for structure noise; when both occur, these assumptions amplify each other’s errors. In this paper, we propose DeFence, a robust node classification model under joint label-structure noise. DeFence constructs anchors from pre-propagation embeddings, avoiding reliance on noisy edges and labels. These anchors regularize representation learning and predictions under joint noise, while the resulting embeddings are periodically reclustered to refine the anchors. This mutual refinement is realized through anchor-filtered contrastive learning, anchor-posterior alignment, and gradient-matched supervision. We theoretically show that, under a better-than-random anchor assumption, anchor filtering improves class discrimination. Experiments on standard benchmarks demonstrate consistent improvements over strong baselines.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
Robust GNN, Noise GNN, Graph Neural Network

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_g78cOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper studies graph robustness against joint lable-structure noise under node classification setting. Specifically, the authors notice that if both label and structure are noisy, feature can provide clean information. So, they firstly construct anchors from feature, then enhance the similarity between two end-nodes of one edge if the nodes belong to the same anchor, and also align the supervised gradient with that of high-confidence unlabeled nodes.

Critical strengths

None.

Critical weaknesses
  1. Motivation. This paper mainly claims that previous models focus on either label noise or structure noise, but none of them solves the both, so the authors did. It sounds like a A+B story, not very interesting. Among label, structure and feature, there are totally 3 binary combinations, and 1 ternary combination. Therefore, there could be 4 such papers. Also, the proposed technique mainly relies on clean feature, but if feature is also noisy, also a common setting, this method will fail.

  2. Method. The key of DEFENCE is to construct a high-quality set of anchors, Eq. (1) and (2). This construction is composed by one-layer MLP, K-Means, and label alignment. I concern if the one-layer MLP + K-Means is strong enough to build anchors. Given the lables are noisy, I also concern if the label alignment and hard anchor assigment can improve the anchors.

  3. Theory. In theory, the authors attempt to prove if rir_i is better than random guess, LalignL_{align} will correct more on the nodes with wrong labels.

  • The conclusion is straight. If we assume rir_i is better than random guess, then rir_i contains some ground-truth label information. Therefore, if one node is given an incorrect class, it deviates more from rir_i, and naturally has a larger gradient if we use KL(ri∣∣p)KL(r_i||p)
  • The reason for Eq. (4) LalignL_{align} is to reduce the drift induced by corrupted labels and noisy propagation. So, the theory needs to prove the gradient alignment is better than no alignment.
  • The assumption rir_i is better than random guess is not always true, especially for one-layer MLP + K-Means.
  1. Experiments. The authors also need to show that with the increase of noise level, the proposed DEFENCE performance decreases much slower than baselines, which indicates the method is robust.
Questions for authors

Please check the weaknesses part.

2 comments

We thank the reviewer for the detailed comments. We clarify the contribution and provide additional evidence below. We have also incorporated the corresponding revisions into the manuscript, with all changes highlighted in blue.

W1. Novel research problem

Motivation and methodological contribution. The joint label–structure noise problem has already been studied by RNCGLN and GDF. We already discussed both methods in the Related Work section and included them in our experimental comparisons. Our contribution is an anchor-guided learning strategy that addresses the mutual reinforcement of label and structure errors: corrupted edges distort neighborhood aggregation, while noisy supervision reinforces the resulting representation errors.

DeFence addresses this interaction through a periodically fixed anchor reference constructed from pre-propagation feature embeddings. The reference guides contrastive pair selection and posterior regularization. Under the respective assumptions of Theorems 1 and 2, even anchors that are only better than random increase positive- and negative-pair purity and induce stronger expected output-level correction for predictions following corrupted labels. These effects support robust learning by providing cleaner pairwise supervision and selectively regularizing predictions driven by noisy labels. Our additional anchor-quality and pair-purity measurements provide empirical support for this mechanism under joint noise (see the answers for W2 and W3).

Scope and future work. This work focuses on joint label and structure corruption; feature corruption lies outside the evaluated scope. Extending DeFence to simultaneous feature, label, and structure corruption is an important direction for future work.

We have clarified our methodological contribution in the Introduction by explaining how periodically fixed anchors from pre-propagation feature embeddings guide contrastive pair selection and posterior regularization under joint label-structure noise. We have also highlighted robustness to simultaneous feature, label, and structure corruption as an important direction for future work in the Conclusion Section.

W2. Anchor reliability through training

Anchor reliability throughout training. To directly assess the anchors produced by the one-layer encoder and K-means, we compare anchor and model accuracy at 10%, 30%, and 100% of training across three datasets and two joint-noise settings.

Anchors are more accurate than model predictions early in training in all six settings. At 10% of training, anchor accuracy exceeds the corresponding model performance in every setting. Anchors remain more accurate in four of six settings at 30%, while model predictions surpass anchors in all six settings by the end of training. This pattern supports the intended role of anchors as informative early guidance, with final predictions exceeding the accuracy of that guidance.

Dataset(Label noise, Structure noise)Total epochs10%: Anchor / Model (%)30%: Anchor / Model (%)100%: Anchor / Model (%)
Cora(0.1, 0.1)42945.05 / 7.8758.57 / 73.3663.51 / 79.08
(0.2, 0.2)29928.12 / 7.6739.29 / 8.1058.45 / 72.60
PubMed(0.1, 0.1)27163.65 / 39.2769.79 / 70.6875.75 / 76.98
(0.2, 0.2)16765.18 / 39.9671.27 / 68.2969.15 / 70.23
Citeseer(0.1, 0.1)10241.21 / 33.7870.87 / 47.8672.26 / 83.06
(0.2, 0.2)11739.19 / 13.3760.44 / 51.5971.47 / 78.83

We added this analysis in Appendix A9.

W3. Anchor reliability and Theorem.

Our analysis explains how imperfect but informative anchors support robust learning through two distinct mechanisms.

Theorem 1 — Cleaner pairwise supervision. Theorem 1 quantifies how better-than-random hard assignments improve positive- and negative-pair purity in L_cont. This reduces cross-class positive pairs and same-class negative pairs, limiting erroneous attraction and repulsion between node representations. The resulting cleaner contrastive supervision reduces the influence of corrupted graph structure on representation learning.

Theorem 2 — Selective regularization. Under the classwise better-than-random anchor assumption, Theorem 2 shows that L_align applies stronger output-level correction when predictions follow incorrect labels than when they follow clean labels. As shown in our response to W2, anchors outperform model predictions in the early stage of training. At this stage, L_align guides predictions toward a more accurate feature-derived reference. Its stronger correction for predictions following incorrect labels opposes noisy-label fitting, while its weaker correction for predictions following clean labels limits interference with reliable supervision, supporting robust learning under joint noise.

Empirical support. We evaluate whether the constructed anchors satisfy the informativeness conditions required by Theorems 1 and 2 and provide reliable pairwise supervision under joint noise. Hard-anchor accuracy exceeds the corresponding chance level in all 15 evaluated settings (Table 1). For Theorem 2, the minimum classwise mean true-class anchor probability exceeds (1/C) in all 10 evaluated settings (Table 2). Anchor filtering also improves both positive- and negative-pair purity on all three datasets (Table 3). These results support the anchor-informativeness premise and directly corroborate the pair-filtering benefit.

Table 1

Dataset(0.1, 0.1)(0.1, 0.2)(0.1, 0.5)(0.2, 0.1)(0.5, 0.1)Random guessing (1/C, %)
Cora59.0057.5944.2461.6443.4214.29
PubMed71.0671.5646.0472.0973.5933.33
Citeseer77.5973.5253.9076.6069.7916.67

Table 2

Dataset(Label noise, Structure noise)Minimum classwise mean true-class anchor probability (%)Random guessing (1/C, %)
PubMed(0.1, 0.1)36.6133.33
(0.1, 0.2)38.6233.33
(0.1, 0.5)37.9833.33
(0.2, 0.1)37.2933.33
(0.5, 0.1)33.6233.33
Citeseer(0.1, 0.1)33.4316.67
(0.1, 0.2)35.3516.67
(0.1, 0.5)41.4916.67
(0.2, 0.1)33.4416.67
(0.5, 0.1)18.8016.67

Table 3

DatasetPositive purity (before → after)Negative purity (before → after)
Cora64.76 → 80.1082.27 → 87.57
PubMed68.75 → 79.0164.37 → 74.77
Citeseer74.92 → 93.0683.04 → 90.95

We added this analysis in Appendix A9.

W4. Performance degradation as noise levels increase

Our existing experiments evaluate noise rates of 0.1 and 0.2 in Table 1 and extend the range to 0.5 in Appendix A.5 (Table 5). We quantify degradation by computing the accuracy drop from 0.1 to 0.5 while keeping the other noise rate fixed at 0.1.

Averaged over all four datasets, DeFence’s accuracy drops by 18.93 percentage points under increasing label noise and 4.40 points under increasing structure noise, compared with 24.75 and 6.58 points for GDF, respectively. For increasing label noise on Cora and Citeseer, the mean drop is 14.30 points for DeFence versus 23.50 points for RNCGLN; the other datasets are excluded from this comparison because RNCGLN encounters out-of-memory errors.

The degradation trend depends on the dataset, baseline, and noise source. RNCGLN has smaller structure-noise drops on the two shared datasets, while DeFence achieves higher accuracy at the high-noise endpoint.

We have added these comparisons of accuracy drops to Section 4.2 of the revised manuscript.

Reviewer_P4XWOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces DeFence, a robust node classification method under joint label and structure noise. The method constructs feature-derived anchors through K-means clustering, using them to guide representation learning and reduce the influence of corrupted edges and labels. The method contains three components: 1) anchor-filtered contrastive learning to improve node representations, 2) a posterior alignment loss to regularise predictions, and 3) gradient-matched supervision to reduce influence of mislabelled instances. There were concerns on the method and evaluations.

Critical strengths
  1. The paper targets the problem of node classification under joint label and structure noise, an important setting in robust graph learning.

  2. Using feature-based anchors to guide learning under noisy labels and graph structures is reasonably motivated.

  3. DeFence demonstrates improvement performance across multiple benchmarks and existing methods, and provides several ablations, theoretical analysis to support the efficacy.

Critical weaknesses
  1. The motivation and reliability of the constructured anchors appears to be insufficiently justified. Line 130 states that the anchors are constructured from pre-propagation embeddings without directly using noisy edges or labels, however, these embeddings are still trained using objectives influenced by both noise sources. Furthermore, the cluster to class alignment also relies on potentially corrupted training labels. Therefore, this raises concerns about whether the anchors are genuinely providing reliable guidance or not, particularly when errors may accumulate during iterative refinement.

  2. Lines 304-322 suggests that the theorems assumes the anchors are more accurate than random guessing. This assumption may be too strong given that paper does not sufficiently investigate whether this holds during training, particularly under severe noise or when node features are weakly informative.

  3. Based on the methods description in Section 3, DeFence seems to rely heavily on the assumption that feature-based clustering can recover meaningful class information. However, the experiments do not directly examine how performance changes when features are less informative, clusters overlaps, or anchor assignments are incorrect. This can be important given that incorrect anchors could reinforce errors through both contrastive learning and posterior alignment.

  4. The evaluation setting described in Section 4 primarily uses synthetic label corruption and randomly added or removed edges on four standard graph datasets. These perturbations may not capture realistic, targeted, or correlated label-structure corruption. More challenging settings such as real-world graph shifts (e.g., multi-graph scenarios) could be benefitical over purely controlled noise assumptions.

Questions for authors
  1. Section 3 uses observed training labels to align clusters with classes. How sensitive is this alignment to label corruption (e.g., what happens at higher noise rates) ?

  2. Could the authors further clarify the individual contributions of the three components, particularly given the relatively small improvements observed in some ablation settings (e.g., Table 2 shows performance drop with w/o L_cont and w/o L_align are relatively marginal)?

  3. Would directly applying feature-based clustering or pseudo-label regularisation to a decoupled GNN backbone achieve similar improvements?

3 comments

We thank the reviewer for the positive comments. We clarify the concerns below and have incorporated the corresponding revisions into the manuscript, with all changes highlighted in blue.

W1 / Q1. Anchor reliability and iterative refinement

The role of pre-propagation anchors is to provide a periodically fixed, feature-derived reference for learning under joint noise. Their usefulness depends on whether they retain informative class structure after training with the corrupted data. We therefore directly evaluate anchor and model accuracy at three stages of training (Table 1).

Table 1

Dataset(Label noise, Structure noise)Total epochs10%: Anchor / Model (%)30%: Anchor / Model (%)100%: Anchor / Model (%)
Cora(0.1, 0.1)42945.05 / 7.8758.57 / 73.3663.51 / 79.08
(0.2, 0.2)29928.12 / 7.6739.29 / 8.1058.45 / 72.60
PubMed(0.1, 0.1)27163.65 / 39.2769.79 / 70.6875.75 / 76.98
(0.2, 0.2)16765.18 / 39.9671.27 / 68.2969.15 / 70.23
Citeseer(0.1, 0.1)10241.21 / 33.7870.87 / 47.8672.26 / 83.06
(0.2, 0.2)11739.19 / 13.3760.44 / 51.5971.47 / 78.83

Anchors outperform model predictions in the early stage in all six evaluated settings. At the intermediate stage, anchors remain more accurate in four settings, while model predictions surpass anchors in all six settings by the end of training. These observations support the intended role of anchors as useful early guidance, with the learned model subsequently exceeding the accuracy of its reference. We added this analysis in Appendix A9.

Additionally, our theoretical analysis explains why anchors need not be highly accurate to provide useful regularization. Theorem 1 shows that better-than-random hard assignments improve positive- and negative-pair purity in L_cont. Under its classwise better-than-random condition, Theorem 2 establishes stronger expected output-level correction by L_align for predictions following incorrect labels at matched confidence. These results motivate the use of informative but imperfect anchors as learning references. Please see our response to W2 for empirical assessment of anchor informativeness under increasing noise, and our response to W3 for the pair-purity evidence and discussion of imperfect anchors.

W2. Empirical assessment of anchor informativeness

We separately assess the anchor informativeness required by Theorems 1 and 2. Theorem 1 requires the probability of a correct hard anchor assignment to exceed 1/C where C is the number of classes, while Theorem 2 requires the mean soft anchor probability assigned to the true class to exceed 1/C for every class.

Hard-anchor informativeness (Theorem 1). Anchor accuracy exceeds the corresponding chance level in all 15 evaluated settings, including severe corruption where either the label or structure noise ratio reaches 0.5 (Table 2).

Table 2

Dataset(0.1, 0.1)(0.1, 0.2)(0.1, 0.5)(0.2, 0.1)(0.5, 0.1)Random guessing (1/C, %)
Cora59.0057.5944.2461.6443.4214.29
PubMed71.0671.5646.0472.0973.5933.33
Citeseer77.5973.5253.9076.6069.7916.67

Classwise soft-anchor informativeness. (Theorem 2) To assess Theorem 2’s condition directly, we average the true-class anchor probability within each ground-truth class and report the minimum across classes. This minimum exceeds 1/C in all ten evaluated settings, supporting the classwise informativeness premise even under severe label or structure corruption (Table 3).

Table 3

Dataset(Label noise, Structure noise)Minimum classwise mean true-class anchor probability (%)Random guessing (1/C, %)
PubMed(0.1, 0.1)36.6133.33
(0.1, 0.2)38.6233.33
(0.1, 0.5)37.9833.33
(0.2, 0.1)37.2933.33
(0.5, 0.1)33.6233.33
Citeseer(0.1, 0.1)33.4316.67
(0.1, 0.2)35.3516.67
(0.1, 0.5)41.4916.67
(0.2, 0.1)33.4416.67
(0.5, 0.1)18.8016.67

These measurements provide direct empirical support for anchor informativeness under the evaluated joint-noise conditions. Feature-dependent limitations and extensions are discussed in our response to W3. We added this analysis in Appendix A9.

W3. Imperfect anchors and feature-dependent limitations

Our analysis concerns informative but imperfect anchors; it does not require error-free assignments or perfectly class-separated clusters. Theorem 1 explains how better-than-random hard assignments improve the pairwise supervision used by L_cont, reducing erroneous attraction between different-class nodes and repulsion between same-class nodes. Theorem 2 characterizes the selective action of L_align which regularizes predictions toward the soft anchor distribution and applies stronger expected output-level correction to predictions following incorrect labels at matched confidence.

We directly examine whether imperfect anchors yield useful pairwise information. Anchor filtering improves both positive- and negative-pair purity on all three datasets:

Table 4

DatasetPositive purity (before → after)Negative purity (before → after)
Cora64.76 → 80.1082.27 → 87.57
PubMed68.75 → 79.0164.37 → 74.77
Citeseer74.92 → 93.0683.04 → 90.95

We added this analysis in Appendix A9.

This work focuses on joint label and structure corruption. Systematically weakening feature information, increasing cluster overlap, and introducing controlled anchor-assignment errors would further characterize the limits of anchor-guided learning. We have identified these extensions, including explicit feature corruption, as directions for future work in the Conclusion Section.

W4. Realistic label-structure corruption

We compare DeFence with two methods designed for joint label–structure noise: GDF and RNCGLN. In the coupled setting, mislabeled training nodes connect to nodes whose clean class matches their observed label. DeFence exceeds both methods on all three datasets (Table 5). In the DICE-inspired setting [1], cross-class edge additions are combined with noisy labels; DeFence exceeds GDF and RNCGLN on all three datasets (Table 6).

Table 5

DatasetDeFence (proposed)GDFRNCGLN
Cora89.0086.7087.40
PubMed85.7081.7082.10
Citeseer92.3589.5774.26

Table 6

DatasetDeFence (proposed)GDFRNCGLN
Cora77.1070.7071.00
PubMed74.5066.9074.50
Citeseer83.1980.4774.14

We added this analysis in Appendix A11. Real-world graph shifts and multi-graph evaluation are future work and included in the Conclusion Section.

Q2: Contributions of the three components

Table 2 of the manuscript already isolates the contribution of each component. The contrastive and posterior-alignment components are implemented through their respective loss terms: removing L_cont disables anchor-filtered contrastive regularization, and removing L_align disables posterior–anchor regularization. The w/o-grad variant replaces gradient-matched supervision with uniformly weighted cross-entropy. Each variant therefore disables one learning mechanism while retaining the other two.

The resulting differences measure each component’s incremental contribution within the complete framework, and their magnitudes vary across settings. For example, removing contrastive regularization reduces Cora accuracy from 72.6% to 67.6% at (0.2,0.2). On Flickr at (0.2,0.1), removing the contrastive, alignment, and gradient components reduces accuracy from 71.5% to 69.7%, 69.9%, and 69.9%, respectively.

Thus, the smaller drops in some settings should be interpreted alongside these larger effects: Table 2 quantifies the setting-dependent contribution of each mechanism when the remaining robustness mechanisms are still active.

We have clarified in Section 4.4 how each ablation disables the corresponding learning mechanism.

Q3: Other regularizations

We compare DeFence with two simpler regularization variants using the same decoupled GNN backbone on Cora. Cluster KL aligns predictions with soft assignments from feature clustering, while Prediction KL aligns predictions with cached model predictions refreshed every ten epochs.

DeFence outperforms both variants in all four evaluated settings, exceeding the stronger variant by 2.30–3.70 percentage points (Table 7). These results show that directly applying clustering or prediction-based regularization alone does not reproduce DeFence’s improvements.

Table 7

(Label noise, Structure noise)DeFence (proposed)Cluster KLPrediction KL
(0.1, 0.1)79.0074.9075.30
(0.1, 0.2)75.4072.1073.10
(0.2, 0.1)73.7069.5070.50
(0.2, 0.2)72.6069.7067.90

We added this analysis in Appendix A10.

References.

[1] Adversarial attacks on graph neural networks via meta learning, ICLR 2019

The response addressed all my concerns adaquetly. I would like to maintain my score, all the best to the authors.

Reviewer_K4eMOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates robust node classification under joint label and graph structure corruption. The authors propose DeFence, a framework that constructs feature-derived anchors from node embeddings before graph propagation to provide a reference less affected by structural noise. The method combines anchor-filtered contrastive learning, posterior-anchor alignment, and gradient-matched supervision.

Experiments on four graph benchmarks show competitive performance under varying levels of label and structural corruption. The problem is relevant, and the reported improvements are encouraging. However, the reliability of the constructed anchors under noisy features and labels remains insufficiently examined. The distinction from existing robust graph learning methods is also unclear, and the experiments do not fully isolate the contributions of the proposed components. These issues limit the evidence for the method's robustness and technical advantages.

Critical strengths
  1. The paper addresses a relevant problem where label and structural noise occur simultaneously. The motivation is reasonable, as these two types of corruption can interact and undermine conventional graph learning methods.

  2. DeFence presents a coherent framework that uses feature-derived anchors to guide structural regularization and noisy-label supervision. The integration of anchor filtering, posterior alignment, and gradient-based weighting is technically reasonable.

  3. The method achieves competitive results across four graph benchmarks under various noise configurations, with notable improvements on Flickr. The additional experiments on stronger corruption, alternative backbones, and runtime provide useful evidence of its effectiveness.

Critical weaknesses
  1. The reliability of the constructed anchors is not sufficiently justified. Although anchors are derived from pre-propagation embeddings, the feature encoder is still trained using noisy labels and graph structure. Moreover, Hungarian matching explicitly uses corrupted training labels to assign clusters to classes. The paper lacks direct evidence that these anchors remain reliable under increasing noise levels.

  2. The performance gains are not clearly isolated from the backbone and existing regularization techniques. DeFence builds on APPNP but does not include standard APPNP as a standalone baseline. The ablations also do not fully separate the effects of feature-only representations, clustering regularization, and anchor refinement. Comparisons with APPNP, MLP, and simpler clustering-based variants are needed to establish the contribution of the proposed design.

  3. The theoretical analysis relies on assumptions that are not fully established. Theorem 1 appears to assume independence between node-level anchor errors when computing pairwise agreement probabilities, which does not follow from the stated marginal error model. This assumption is questionable for correlated graph representations. Theorem 2 only characterizes gradient magnitudes under specific conditions and does not establish improved robustness during training.

  4. The contribution of gradient-matched supervision remains uncertain. The reference gradient is estimated from high-confidence predictions of a model already trained on corrupted data, making its reliability questionable. The ablations show limited or inconsistent gains from this component. A comparison with simpler confidence-based or loss-based weighting, together with the weight distributions of clean and corrupted samples, is needed.

Questions for authors
  1. How does anchor quality change with label and structural noise? In particular, how reliable is the Hungarian cluster-to-class matching when training labels are corrupted?

  2. How does DeFence compare with standard APPNP and simpler clustering-based regularization under identical training settings? Can the gains from anchor refinement be isolated from those of the backbone?

  3. Does Theorem 1 require independence of anchor assignment errors across nodes? If so, how is this assumption justified for graph data, where neighboring node representations and errors may be correlated?

3 comments

We thank the reviewer for the constructive feedback. We address each concern below and have incorporated the corresponding revisions into the manuscript, with all changes highlighted in blue.

W1 / Q1. Anchor reliability

Our theoretical analysis explains why anchors need not be highly accurate to provide useful regularization. According to Theorems 1 and 2, better-than-random anchors suffice to improve positive- and negative-pair purity and to induce stronger expected output-level correction for predictions following incorrect labels.

We therefore examine whether the learned anchors retain the required informativeness under joint noise. Hard-anchor accuracy exceeds the corresponding chance level in all 15 evaluated settings (Table 1). Separately, the minimum classwise mean true-class anchor probability exceeds (1/C) in all 10 settings evaluated for Theorem 2 (Table 2). Anchor filtering also improves both positive- and negative-pair purity on all three datasets (Table 3). These results support the informativeness premise of our analysis and directly corroborate the pair-filtering benefit.

Table 1

Dataset(0.1, 0.1)(0.1, 0.2)(0.1, 0.5)(0.2, 0.1)(0.5, 0.1)Random guessing (1/C, %)
Cora59.0057.5944.2461.6443.4214.29
PubMed71.0671.5646.0472.0973.5933.33
Citeseer77.5973.5253.9076.6069.7916.67

Table 2

Dataset(Label noise, Structure noise)Minimum classwise mean true-class anchor probability (%)Random guessing (1/C, %)
PubMed(0.1, 0.1)36.6133.33
(0.1, 0.2)38.6233.33
(0.1, 0.5)37.9833.33
(0.2, 0.1)37.2933.33
(0.5, 0.1)33.6233.33
Citeseer(0.1, 0.1)33.4316.67
(0.1, 0.2)35.3516.67
(0.1, 0.5)41.4916.67
(0.2, 0.1)33.4416.67
(0.5, 0.1)18.8016.67

Table 3

DatasetPositive purity (before → after)Negative purity (before → after)
Cora64.76 → 80.1082.27 → 87.57
PubMed68.75 → 79.0164.37 → 74.77
Citeseer74.92 → 93.0683.04 → 90.95

We added this analysis in Appendix A9.

W2 / Q2. Backbone and simpler regularization

We compare two baseline models (MLP and APPNP) and two regularizers (Cluster KL and Prediction KL) on Cora. Cluster KL aligns APPNP predictions with soft assignments from feature clustering; Prediction KL aligns them with cached model predictions refreshed every 10 epochs. DeFence outperforms all baselines, supporting gains beyond the backbone and simple regularization (Table 4).

Table 4

(label, edge)DeFence (proposed)MLPAPPNPCluster KLPrediction KL
(0.1, 0.1)79.0054.5075.2074.9075.30
(0.1, 0.2)75.4054.5073.5072.1073.10
(0.2, 0.1)73.7052.9068.6069.5070.50
(0.2, 0.2)72.6052.9068.5069.7067.90

We added this analysis in Appendix A10.

W3 / Q3. Assumptions in the theoretical analysis

Theorem 1 — Pair reliability. Independence is an analytical simplification used to derive an explicit quantitative bound. We further evaluate filtering with the actual learned anchors under five (label, structure) noise settings: (0.1, 0.1), (0.1, 0.2), (0.1, 0.5), (0.2, 0.1), and (0.5, 0.1) (see Table 3 of our answer to W1). The results demonstrate that the pair-filtering benefit is observed with learned anchors whose errors may be correlated, beyond the independent-error model used for the explicit bound.

Theorem 2 — Selective correction. Theorem 2 characterizes how posterior–anchor alignment responds to predictions following clean versus corrupted labels at the same prediction confidence. Under its classwise better-than-random anchor assumption, the alignment loss produces a larger expected squared gradient with respect to the output logits for predictions following incorrect labels, averaging over incorrect classes. This establishes the selective action of the regularizer at the output level. The classification and ablation experiments in Section 4 separately assess its contribution to end-to-end robustness.

We additionally examine whether the anchors satisfy the informativeness condition underlying this analysis. Theorem 2 requires the mean soft anchor probability assigned to the true class to exceed 1/C for every class. Accordingly, we compute this mean within each ground-truth class and report the minimum across classes (see Table 2 in our answer to W1).

The reported minimum exceeds 1/C in all 10 evaluated settings. Thus, every evaluated class satisfies the empirical counterpart of Theorem 2’s anchor-informativeness condition, including under severe label or structure corruption.

We added this discussion in Appendix A9.

W4. Comparison with other sample-weighting methods

We compare DeFence with confidence-based and exponential-loss-based sample weighting on Citeseer. Confidence uses the maximum class probability, wi=max⁡cpi(c)w_i=\max_c p_i(c); exp(-CE) uses the observed-label probability, wi=pi(y~i)w_i=p_i(\tilde y_i). The results show higher accuracy than both alternatives (Table 5). We also included a weight-distribution figure in the revised paper comparing clean and corrupted training nodes. As a result, clean nodes receive higher weights on average (Figure 6 in Appendix A12).

Table 5

(label, edge)DeFence (proposed)Confidenceexp(-CE)
(0.1, 0.1)83.1082.3781.07
(0.1, 0.2)79.7079.6079.65
(0.2, 0.1)81.7080.7480.16
(0.2, 0.2)78.8077.0577.41

We added this discussion in Appendix A12.

While I appreciate the clarifications provided and the effort put into the rebuttal, I have carefully considered the authors' feedback and the proposed revisions. However, I still maintain my original score of 2. I hope the authors can further strengthen this work in the future.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.