acceptodds
Under review as a conference paper at ICLR 2027

Uncertainty is not Enough: Misclassification Detection via Distribution-aware Bayesian Graph Neural Networks

Abstract

Bayesian Graph Neural Networks (BGNNs) have achieved significant accomplishments in uncertainty quantification for graph-structured data, particularly in isolating epistemic uncertainty. However, traditional BGNNs primarily parameterize uncertainty over model weights and posterior predictive distributions, which leads to miscalibration under structural and attribute noise. For node classification tasks, this challenge is often noticed as confident misclassifications – underestimating prediction uncertainty on out-of-distribution (OOD) samples or under adversarial perturbations – and results in severe risks for downstream applications such as medical diagnostics and threat detection. To address this challenge, we propose Distribution-Aware Bayesian Graph Neural Networks (DA-BGNNs), a novel framework designed for robust node-level classification. DA-BGNN leverages Monte Carlo sampling alongside a distributional shift metric that explicitly measures the discrepancy between correctly classified and misclassified node representations. By combining both the post-training variational posterior and local neighborhood feature distributions, while considering localized heterophily via variational inference, our framework effectively pinpoints misclassifications produced by adversarial perturbations and topological noise. Our method demonstrated state-of-the-art (SOTA) results that consistently outperform conventional non-Bayesian GNN methods and standard BGNN baselines. Under both in-distribution and adversarial perturbation, DA-BGNN indicates an average improvement of 16% in AUPRC and reduces Expected Calibration Error (ECE) by 25% for node classification tasks. Furthermore, it enables more reliable and faithful uncertainty estimation across predictive entropy, variance, and distribution shift.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Uncertainty Quantification, Epistemic Uncertainty, Bayesian Representation Learning, Graph Neural Networks, Conformal Prediction, Post-hoc Calibration, Adversarial Attacks, Node Classification, Misclassification Detection, Probability Distribution, Variational Inference, Scientific Machine Learning, Wasserstein Distance, Weisfeiler-Lehman Graph Kernal

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_BeZ8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper studies misclassification detection for semi-supervised node classification. It argues that prior methods primarily parameterize uncertainty over model weights and posterior predictive distributions, which is insufficient when errors arise in heterophilic neighborhoods. The proposed DA-BGNN augments a variational Bayesian GCN with a composite score that combines Monte Carlo predictive entropy, epistemic variance, a node-level heterophily ratio αi\alpha_i, and a representation-shift term δi\delta_i, implemented as a normalized sum. Experiments on Cora, CiteSeer, PubMed, and Amazon Computers show the best AUROC/AUPRC for misclassification detection (Table 2) while remaining competitive in accuracy and ECE.

Critical strengths
  1. Confident misclassification of GNNs is a real risk for downstream applications such as fraud detection and medical diagnosis. Improving misclassification detection on node classification tasks is a well-motivated direction.

  2. The empirical motivation is clear: neighborhood entropy is 3–11× higher in high-heterophily regions (Table 8, Fig. 1). The ablation (Table 3) isolates heterophily as the dominant contributor (AUPRC 0.445 → 0.522).

Critical weaknesses

W1. The paper reports the AUPRC gain as about 16% in the abstract, 21.0% / 18.8% in Table 4, and about 34% in Section 5 (vs. BGNN in Table 2). The ECE reduction is reported as 25% in the abstract and 19.6% in Table 4, while Table 2 implies roughly 62% relative to GCN (0.099 → 0.038). Appendix K.4 acknowledges that the percentages "depend on the chosen baseline." The authors should state one clearly defined headline comparison (fixed baseline, fixed setting) and use it consistently across the abstract, main text, and tables.

W2. Presentation needs substantial improvement. (i) Broken cross-references ("Equation ??") appear twice on page 4. (ii) There are numerous typos, e.g., "appraoch", "seperates", "EXPERIEMNTAL", "Laslty", "catach", "SOT results", "replied on". (iii) The introduction is written informally ("Why?", "There is one more question to ask ourself"). (iv) The text states that "All proofs are supplied in Appendix," but the paper contains no theorem or proof. (v) Appendix D contains a textbook-style numeric forward-pass example that does not relate to the contributions. (vi) The Ethics and Reproducibility statements are empty.

Questions for authors

Q1. Discrepancy between Table 2 and Table 3. The full DA-BGNN reports 0.880 / 0.571 (AUROC / AUPRC) in Table 2 but 0.854 / 0.510 in Table 3. Both captions state an average over the same four datasets. Averaging the per-dataset results in Table 5 reproduces Table 3, but not Table 2; the same mismatch appears for the GCN and BGNN rows. Could the authors clarify which setting (splits, perturbation, score variant) Table 2 reports? Relatedly, Table 3 suggests that adding the representation-shift term δ\delta slightly reduces performance compared with "+ heterophily α\alpha" (0.861 / 0.522). What is the justification for including δ\delta in the final score?

Q2. Computation of αi\alpha_i may cause possible label leakage. Eq. 50 defines αi\alpha_i with the node's own label yiy_i, which is unavailable for test nodes. Section 3.1 states that heterophily statistics use only training and validation labels. Please confirm that no test labels are used at any stage. This matters because α\alpha is the dominant source of improvement in Table 3.

2 comments
AuthorsOpenReview ↗

Rebuttal (p.1)

Dear Reviewer BeZ8,

We sincerely thank you for your careful reading and for recognizing that confident GNN misclassification is a real risk, as well as for noting that our ablation isolates heterophily as the dominant contributor. Your rating is clear and your concerns are concrete, we treat them seriously. We have carefully revised the manuscript accordingly. Please refer to the updated PDF.

Below we provide point-by-point responses to all weaknesses and questions.

W1: (abstract / Table 4 / Section 5 / ECE claims)

We agree that the submitted draft mixed baselines and settings with gains. The revision has adopted one headline comparison and uses it consistently:

DA-BGNN vs BGNN (MC), clean in-distribution test nodes, mean over Cora, CiteSeer, PubMed, Amazon Computers; identical splits, main-table (calibrated / validation-tuned) detector.

Under that definition, AUPRC goes 54.8 → 83.7 (≈ 53% relative), AUROC 85.6 → 97.3, while Acc/ECE stay at the frozen BGNN values (81.1 / 0.029). We no longer promote ECE “reduction by DA-BGNN”, because the method is post-hoc and does not change predictions/calibration of the backbone.

Appendix shows percentages that“depend on the chosen baseline”. Please read the revised PDF:

  1. Abstract / main results prose use only the headline above (≈ 53% AUPRC vs BGNN).
  2. Secondary comparisons (e.g., vs NAPS AUPRC 71.0 → 83.7, ≈ 18% relative) are labeled explicitly when mentioned and are never swapped into the abstract.
  3. Ablation decimals (0.445 → 0.522 for +α+\alpha) remain absolute table values, not re-packaged as the abstract percentage.
  4. ECE: NAPS still has the best ECE (0.015); DA-BGNN matches BGNN (0.029). Claiming a large ECE cut vs GCN as a DA-BGNN contribution was misleading for a post-hoc calibration, we removed that.

Main table:

MethodAccECEAUROCAUPRC
BGNN (MC)81.10.02985.654.8
NAPS79.70.01592.671.0
DA-BGNN81.10.02997.383.7

W2 Presentation

We agree presentation was below standard in the submitted draft. The revision (i) fixes broken equation references, (ii) corrects the listed typos, (iii) tones down informal intro phrasing, (iv) makes Theorems 1–3 / Lemma 1 and their proofs explicit in the appendix (so “All proofs are in the Appendix” is no longer vacuous), (v) shortens / relocates the textbook forward-pass example so it no longer distracts from the contribution, and (vi) fills Reproducibility (and Ethics, as required) with concrete pointers to splits, seeds, Algorithm 1, and anonymized code.

(i) Broken cross-references (“Equation ??”).
Caused by duplicate / missing equation labels in the draft. The revision keeps a single \label{eq:bayesian} (and related labels) and verified references on page 4 resolve.

(ii) Typos (appraoch, seperates, EXPERIEMNTAL, Laslty, catach, SOT results, replied on, etc.).
All listed instances and nearby paste artifacts are corrected in the revised PDF. We additionally proofread captions and appendix tables for similar errors.

(iii) Informal introduction.
We rewrite rhetorical asides (“Why?”, “ask ourself”) into standard academic prose stating the research questions and contributions directly.

(iv) “All proofs are supplied in Appendix” with no theorem.
Fair criticism of the submitted draft. The revised appendix now contains:

  • Theorem 1 (heterophily–uncertainty coupling),
  • Lemma 1 (ranking),
  • Theorem 2 (hybrid vs entropy-only Fisher/MSE),
  • Theorem 3 (finite-sample refinement),
    with proofs under the corresponding appendix subsections. The main-text pointer now matches actual content.

(v) Appendix D textbook forward-pass.
We agree it did not support the contribution. In the revision it is removed or reduced to a short pointer so the appendix focuses on Algorithm 1, label-safe α^\hat{\alpha}, ablations, and theory.

(vi) Empty Ethics / Reproducibility statements.
Reproducibility now points to appendix details (datasets/splits, models, Algorithm 1) and the anonymized repository: https://anonymous.4open.science/r/DA-GNN. Ethics is filled with a standard statement that we use public citation/co-purchase benchmarks and release no private personal data, dual-use risk for fraud/medical settings is noted as a motivation for better misclassification detection, not as a claim of clinical deployment.

AuthorsOpenReview ↗

Rebuttal (p.2)

Q1 and Q2

Ablation: mapping 0.880 / 0.571 vs 0.854 / 0.510

Mean over the four datasets:

VariantAUROCAUPRC
Entropy only (HH)0.800.45
BGNN (H+νH{+}\nu)0.800.42
++ heterophily α\alpha0.860.52
Equal-weight full (+,δ+\\,\delta)0.850.51
Validation-tuned full0.880.57

Your concern is exactly what this table shows. The per-dataset equal-weight appendix averages to the equal-weight family (≈ 0.85 / 0.51), not to the tuned row. Heterophily is the main clean equal-weight lift, tuning then moves the full feature set further.

Why keep δ\delta after the equal-weight dip

Under equal weights you are right: +α+\alpha (0.86 / 0.52) beats equal-weight full with δ\delta (0.85 / 0.51). We do not claim that putting weight one on every coordinate is optimal.

We still keep δ=∣zi−c∣2\delta=\\|z_i-c\\|_2 for two reasons.

  1. The score we actually use is not equal-weight. Validation-tuned fusion over (H,ν,α^,π,δ)(H,\nu,\hat\alpha,\pi,\delta) reaches 0.88 / 0.57, which is the decimal counterpart of the story behind the main 97.3 / 83.7 calibrated mean. Reweighting can keep a small weight on δ\delta when it helps a subset of nodes without forcing weight one everywhere.
  2. Under shift, δ\delta helps. With FGSM ε=0.05\varepsilon=0.05, adding δ\delta to H+ν+αH{+}\nu{+}\alpha improves AUPRC on every dataset we report: Cora 0.44→0.46, CiteSeer 0.47→0.49, PubMed 0.37→0.39, Amazon 0.43→0.45. The residual is built for representation shift, that is where it earns its keep.

In other words, heterophily carries the result: δ\delta is justified by tuning and by the perturbation table. Leaving the equal-weight regression in the paper is more honest than deleting δ\delta and pretending the full model is only α\alpha.

GCN / BGNN mismatches

Score variant cannot move GCN. If GCN or BGNN also differed across tables under captions that both claimed the same four-dataset average, that was a caption / aggregation error in the submitted draft—not a second experimental setting. The revision uses one protocol for every baseline row and separates tuned vs equal-weight DA-BGNN in the wording. The DA-BGNN gap you highlighted is explained by the score head, the baseline mismatch is acknowledged and corrected.


Q2. Does computing α\alpha use test labels?

No. This is important precisely because +α+\alpha is the dominant equal-weight gain in Table 3 (about 0.80 / 0.45 → 0.86 / 0.52). If that had used test labels, the ablation would be invalid.

The equation that writes αi\alpha_i with the node’s own label yiy_i is the population definition:

αi⋆=1−1∣N(vi)∣∑j∈N(vi)1[yj=yi].\alpha_i^{\star}=1-\frac1{|\mathcal{N}(v_i)|}\sum_{j\in\mathcal{N}(v_i)}\mathbf{1}[y_j=y_i].

We do not evaluate that formula on VUV_U. At inference we use a label-safe estimator from training and validation labels only, Vobs=VL∪VvalV_{\mathrm{obs}}=V_L\cup V_{\mathrm{val}}. For a test node we look at neighbors that already have observed labels and compare them to the mode of those labels—never to ytesty_{\mathrm{test}}:

\hat{\alpha}_i = 1 - \frac{\left|\left\\{j \in \mathcal{N}_{\mathrm{obs}} : \tilde{y}_j = \mathrm{mode}(\tilde{y}_{\mathcal{N}_{\mathrm{obs}}})\right\\}\right|}{\left|\mathcal{N}_{\mathrm{obs}}\right|},

with a median fallback if there are no observed neighbors. Example: observed neighbor labels 1,1,2\\{1,1,2\\} give α^=1−2/3=1/3\hat\alpha=1-2/3=1/3. The test label is never read.

The same observed-only discipline applies downstream. The heterophily law DHD_H and the percentile πi=F^DH(α^i)\pi_i=\widehat{F}_{D_H}(\hat\alpha_i) are formed from observed nodes; the residual center for δ\delta uses correctly classified observed embeddings. None of α^\hat\alpha, π\pi, δ\delta, or the final score sis_i reads labels from VUV_U at any stage of training, validation tuning, or test scoring.

Section 3.1 already stated that heterophily statistics use training and validation labels; the submitted draft still made it too easy to confuse α⋆\alpha^\star with α^\hat\alpha. The revised PDF places the estimator next to the population equation and spells the procedure in Algorithm 1.

We would appreciate if you could raise your scores accordingly. Thank you.

Best regards,
Authors

Reviewer_71pdOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The potentially original contribution of this paper appears to be a straightforward combination of Bayesian predictive uncertainty, local heterophily, and representation-shift features for detecting node classification errors. The empirical results demonstrate its effectiveness relative to baselines on benchmarks.

Critical strengths

The paper addresses an important reliability problem and motivates incorporating neighborhood structure and distributional information beyond Bayesian predictive uncertainty.

The proposed post-hoc scoring framework combines predictive uncertainty, local heterophily, and representation-shift features without requiring changes to the underlying GCN backbone.

The empirical results demonstrate its effectiveness relative to baselines on benchmarks.

Critical weaknesses
  1. The potentially original contribution of this paper appears to be a straightforward combination of Bayesian predictive uncertainty, local heterophily, and representation-shift features for detecting node classification errors. However, this combination does not yield a new algorithm. The methodology section is only around 1 page. The technical novelty and depth can be enhanced.

  2. The heterophily motivation is also not new. It has been discussed in several papers in the existing literature (e.g., ref. 1).

  3. The central function F is repeatedly introduced through its inputs. The most concrete implementation description is a “normalized sum,” with no complete specification of the normalization and other related context. The manuscript does not establish whether the global heterophily distribution is used to compute a density, a percentile, a standard deviation, or another node-specific quantity.

  4. The manuscript does state that training and validation labels are used exclusively. It does not explain how the definition (31) in Appendix F.3 is converted into an estimator for every test node.

  5. The ablation study in Table 3 seems to favor the simpler method. Heterophily provides a clear gain in this ablation. Adding the representation-shift term reduces both metrics. Adding variance to entropy also does not help in the reported comparison.

  6. The datasets were published mostly ten years ago. The authors may consider more recent and realistic benchmark datasets.

  7. There is a line "8 0.835 0.530" on Line 429 that is not interpretable. There is a question mark?) on Line 204.

  8. No code is provided to support replication of the results.

References

  1. Loveland D, Zhu J, Heimann M, Fish B, Schaub MT, Koutra D. On performance discrepancies across local homophily levels in graph neural networks. In Learning on Graphs Conference 2024 Apr 17 (pp. 6-1). PMLR.
Questions for authors

The authors may address my weaknesses mentioned above.

3 comments
AuthorsOpenReview ↗

Rebuttal (p.1)

Dear Reviewer 71pd,

Thank you for your comments and feedback. We apologize for the initial parts of the submission. We were building an interactive graph tool to visualize node-level heterophily and neighborhood context. Although the interface is still in progress, the detection algorithm and DA-BGNN framework are presented in this paper.

We have revised the manuscript to clarify the scoring function FF, the role of DHD_H, the label-safe estimator α^i\hat\alpha_i, the ablation interpretation, presentation issues, and the anonymized code release. Please refer to the updated paper submission.

W1. Novelty and depth

We understand that BGNN is not a new architecture. The GCN encoder and variational posterior qϕ(θ)q_\phi(\theta) are standard; after training, ϕ\phi is frozen. The contribution is a post-hoc misclassification detector that couples Monte Carlo predictive uncertainty with label-safe local heterophily, a global heterophily percentile, and a representation-shift residual, with a complete inference algorithm and supporting theory. Please refer to the heterophily-aware scoring section, Algorithm 1, and Appendix Theory (Theorems 1–3 and Lemma 1). We have explicitly stated the closed-form of FF in the revised PDF.

W2. Prior work on heterophily

Thank you for pointing out “On performance discrepancies across local homophily levels in graph neural networks,” LoG 2024. We do not claim to discover heterophily as a phenomenon. Prior work analyzes when GNN accuracy varies across local homophily, whereas our method utilizes local heterophily, with its position under DHD_H, as a detection feature for confident errors, jointly with Bayesian predictive uncertainty, through the coordinates (H,ν,α^,π,δ)(H,\nu,\hat\alpha,\pi,\delta) on a frozen BGNN. We have cited Loveland et al. and clarified this complementary material in the revised discussion.

W3. Definition of F

We apologize for introducing FF through its inputs as a “normalized sum,” without specifying the quantity induced by DHD_H. In the revision, DHD_H is used only as an empirical CDF/percentile, not as a density or global standard deviation. After examining α^i\hat\alpha_i from observed labels only (W4), we form DH=1∣Vobs∣∑u∈Vobsδα^uD_H=\frac{1}{|V_{\mathrm{obs}}|}\sum_{u\in V_{\mathrm{obs}}}\delta_{\hat\alpha_u} and set πi=F^DH(α^i)\pi_i=\widehat{F}_{D_H}(\hat\alpha_i). From Monte Carlo samples, we compute entropy HiH_i and variance νiout\nu_i^{\mathrm{out}}. The residual is δi=∣zi−c∣2\delta_i=\\|z_i-c\\|_2, where cc is the mean embedding of correctly classified observed nodes. Each coordinate x∈H,ν,α^,π,δx\in\\{H,\nu,\hat\alpha,\pi,\delta\\} is normalized by Z(x)i=xi−meanVU(x)stdVU(x)+ε\mathrm{Z}(x)_i=\frac{x_i-\mathrm{mean}_{V_U}(x)}{\mathrm{std}_{V_U}(x)+\varepsilon} with ε=10−6\varepsilon=10^{-6}. Then FF is either sihyb=∑kZ(x(k))is_i^{\mathrm{hyb}}=\sum_k\mathrm{Z}(x^{(k)})_i, or a validation-tuned non-negative head si=w⊤uis_i=w^\top u_i. This is exactly Algorithm 1.

W4. Estimator for every test node

The population quantity αi⋆\alpha_i^\star requires labels of ii and its neighbors, therefore it cannot be used directly on unlabeled test nodes. Let Vobs=VL∪VvalV_{\mathrm{obs}}=V_L\cup V_{\mathrm{val}} and let Nobs(vi)\mathcal{N}_{\mathrm{obs}}(v_i) be neighbors with observed labels. If this set is empty, set α^i\hat\alpha_i to the median over VobsV_{\mathrm{obs}}. If vi∈Vobsv_i\in V_{\mathrm{obs}}, compare observed neighbors to the label of ii. If vi∉Vobsv_i\notin V_{\mathrm{obs}}, never read yiy_i. Instead, compare observed neighbors to mode(y~Nobs(vi))\mathrm{mode}(\tilde{y}_{\mathcal{N}_{\mathrm{obs}}(v_i)}). This makes sure every vi∈VUv_i\in V_U uses only labels in VobsV_{\mathrm{obs}} and induces no test-label leakage. Algorithm 1 constructs πi\pi_i and computes the final score.

AuthorsOpenReview ↗

Rebuttal (p.2)

Below we answer W5 to W8 questions directly with the reported tables and ablation studies, all came from the revised PDF.

W5. "The ablation study in Table 3 seems to favor .. comparison."

The reading is correct and there is a reason for that. Heterophily is the largest single gain, but that does not mean the method reduces to α\alpha alone. The main detector is the validation-tuned full score, and δ\delta which helps under perturbation.

Table (mean over Cora / CiteSeer / PubMed / Amazon):

VariantAUROCΔ\DeltaAUROCAUPRCΔ\DeltaAUPRC
Entropy only (HH)0.80—0.45—
BGNN (H+νH{+}\nu)0.800.000.000.42−0.02-0.02
++ heterophily α\alpha0.86+0.07+0.070.52+0.08+0.08
Equal-weight full (++ shift δ\delta)0.85+0.06+0.060.51+0.07+0.07
Validation-tuned full score0.88+0.08\mathbf{+0.08}0.57+0.13\mathbf{+0.13}

What this shows

  • Adding variance to entropy does not help: H+νH{+}\nu leaves AUROC flat and lowers AUPRC. We agree. Posterior variance alone fails to catch many confident errors in mixed neighborhoods—this is why structure is needed.
  • Heterophily provides a clear gain: +α+\alpha is the strongest equal-weight step (+0.07+0.07 AUROC, +0.08+0.08 AUPRC). We agree and highlight this.
  • Equal-weight representation-shift reduces metrics vs +α+\alpha: 0.86/0.52→0.85/0.510.86/0.52 \to 0.85/0.51. We agree: naive weight-1 stacking of δ\delta is not optimal.
  • Why the full method still stands: validation-tuned non-negative weights over (H,ν,α^,π,δ)(H,\nu,\hat\alpha,\pi,\delta) recover 0.88 / 0.57, which beats +α+\alpha alone. Under FGSM (ε=0.05\varepsilon=0.05), H+ν+α+δH{+}\nu{+}\alpha{+}\delta further beats H+ν+αH{+}\nu{+}\alpha on all four datasets (e.g., Cora AUPRC 0.44→0.460.44\to 0.46, CiteSeer 0.47→0.490.47\to 0.49, PubMed 0.37→0.390.37\to 0.39, Amazon 0.43→0.450.43\to 0.45). So δ\delta is most useful under distributional shift, matching its design.

The ablation does not refute DA-BGNN, it localizes where the signal is (+α+\alpha) and shows that the full score needs tuned, rather than equal-weight addition of every term.

W6. "The datasets were published mostly ten years ago..."

Thank you for Reviewer's reminder. We are actively looking for collection of graph datasets, if you happen to come up with sources to some labeled graph datasets. Please do share with us here. Meanwhile, we are unable to find other large scale datasets such as the ones we used for this paper, Cora, CiteSeer, PubMed, and Amazon Computers. This also raise a fair comparison to other UQ sampling papers, for example, the one paper you pointed out and our tested literature, ** BGNN, GPN, NAPS, temperature scaling, and CAGCN etc... which are still commonly reported, as far as we know. Using random datasets and baselines could raise serious flags for publication of the paper, people may argue the method only perform well on certain datasets... Therefore, we use identical splits and confound Acc / NLL / ECE / AUROC / AUPRC comparisons. These datasets are supportive to our claim: Frac. α>0.3\alpha > 0.3 is 0.249 / 0.359 / 0.253 / 0.247, heterophily errors are not vacuous. We could consider more modern datasets for future work.

3. "There is a line "8 0.835 0.530" on Line 429 that is not interpretable."

This has been addressed.

4. "No code is provided to support replication of the results."

An anonymized repository is available at https://anonymous.4open.science/r/DA-GNN

It includes scripts for frozen-BGNN training and the post-hoc DA-BGNN scorer (Monte Carlo moments, label-safe α^\hat\alpha, percentile π\pi, residual δ\delta, equal-weight / validation-tuned fusion).

Thank you again for these feedbacks. We hope this response and revision answers all your questions. Please adjust your scores accordingly.

Best regards,
Authors

Thanks, authors, for the additional clarifications and results. W2 and W3 are addressed.

Regarding W1, I do not mean it is not novel just because it combines existing techniques. My concern is that your combination does not seem nontrivial. Thanks for pointing out Algorithm 1 and the theorems in the appendix. It could be better if the authors summarize the novelty in the algorithm and summarize theoretical results in the main paper, as readers may not read the appendix unless there are references mentioned in the main paper. This weakness is partially addressed.

About W5, the advantage may depend on tuned weights. Did the baselines get comparable tuning?

About W6, I do understand the challenge of finding more recent datasets. But my concern about this still exists.

Thanks for posting the code.

I will increase my rating to weak reject. I will discuss with other reviewers later and may further increase my rating if some of their main concerns are addressed.

Reviewer_iLTeOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

There is a double blind policy violation. I will skip my review.

Critical strengths

N/A

Critical weaknesses

N/A

Questions for authors

N/A

1 comment

Dear Reviewer iLTe,

We sincerely apologize for the initial version of our submission and deeply saddened by the double-blind policy issue. We apologize for this mistake. The submitted PDF contained a non-anonymous GitHub URL that includes one of the authors' nickname. We understand the reason for skipping the reivew. However, we ask the chairs and reviewer to reconsider the updated version for review. The anonymized repository includes the training/evaluation scripts for the BGNN stage and the post-hoc DA-BGNN scorer corresponding to our method. We have replaced the github with an anonymous repository: https://anonymous.4open.science/r/DA-GNN. Please refer to the updated submission.

Due to the original review deferred from technical assessment, we would be grateful if you could kindly reconsider the anonymized revision. We would be happy to clarify any questions about the paper, the code, or the experimental settings.

Thank you for your time and consideration.

Best regards,
Authors

Comments on the paper

1 comment
AuthorsOpenReview ↗

Rebuttal Summary

Dear AC and SAC,

We sincerely appreciate your time and effort in this review process for our submission. We understand that due to the AI usage and busy time of this year, ACs are facing significant workloads, thus we would like to share a brief summary of the rebuttal to assist in your decision-making.

1. We appreciate all reviewers’ recognition of our work’s strengths.

Although the initial reviews were not satisfied, we acknowledged the core contributions and merits of our work, including:

  • (i) The practical importance of the problem

    • BeZ8 - "confident GNN misclassification is a real risk for downstream applications such as fraud detection and medical diagnosis"
    • 71pd - "addresses an important reliability problem"
  • (ii) The clarity of the motivation

    • BeZ8 - "neighborhood entropy is 3–11× higher in high-heterophily regions", "the ablation isolates heterophily as the dominant contributor"
    • 71pd - "motivates incorporating neighborhood structure and distributional information beyond Bayesian predictive uncertainty"
  • (iii) The effectiveness and design of the proposed method

    • 71pd - "empirical results demonstrate its effectiveness relative to baselines on benchmarks", "combines predictive uncertainty, local heterophily, and representation-shift features without requiring changes to the underlying GCN backbone"

2. We wrote a detailed rebuttal which meaningfully improved the quality of our work.

Given by constructive feedback from Reviewers iLTe, 71pd, and BeZ8, we revised several parts of the manuscript to improve readability, fix broken references and presentation issues, and clarify technical formulations. This has significantly improved our work. Furthermore, we added additional detail and clarifications to enable the soundness of the paper, including:

  • Resolution of Policy Issue: We addressed the double-blind policy violation by providing an anonymous repository (https://anonymous.4open.science/r/DA-GNN) containing the training and evaluation scripts for BGNN and the post-hoc DA-BGNN scorer.

  • Clarification of Technical Details: We fully detailed the scoring function FF, empirical percentiles, and the label-safe estimator (α~i\tilde{\alpha}_i) to confirm that no test labels are leaked at any stage of training, validation tuning, or test scoring. Additionally, we made explicit the theoretical proofs (Theorems 1–3 and Lemma 1) in the appendix.

  • Unified Reporting & Ablation Analysis: We unified our headline reporting convention (validation-tuned DA-BGNN vs. standard BGNN on clean in-distribution test nodes), corrected table discrepancies, and provided an in-depth justification for why heterophily drives the clean performance gains while the residual shift δ\delta is tailored for adversarial/distributional shifts.

3. We appreciate all the reviewer's detailed comments. We are still actively responding to any questions that may occur at this time.

We have addressed all concerns, regarding novelty, prior heterophily literature, dataset choices, ablations, and potential label leakage raised by each reviewer. We have implemented updates in the revised manuscript. We are confident that our rebuttal can resolve any technical and presentation concerns.

We hope this summary is helpful in facilitating your decision. Everything in this summary is based on the information from the rebuttal and revised manuscript. Lastly, we thank you and all the reviewers for helping us to improve this work.

With best wishes,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.