acceptodds
Under review as a conference paper at ICLR 2027

On the Generalizability of Non-Causal Subgraphs: Addressing Cue Discordance for Graph OOD Generalization

Abstract

Graph neural networks often exploit spurious correlations in training environments, leading to poor out-of-distribution generalization. Existing graph OOD methods typically seek causal or invariant subgraphs, implicitly treating non-causal structures as poorly generalizable by-products of overfitting. However, in our analysis and experiments, we find that non-causal subgraphs can also exhibit nontrivial cross-sample generalization and provide predictive cues that conflict with those encoded by causal subgraphs. We characterize this phenomenon as cue discordance, where the causal subgraph supports the ground-truth class while the spurious favors a different one. Empirically, we show that cue discordance is widespread and markedly more common among misclassified samples; moreover, increasing the proportion of discordant samples consistently reduces accuracy while increasing classification loss. We further theoretically establish the Discordance-Conditioned Causal Contrast (D3C) criterion, proving that under cue discordance, causal and non-causal subgraphs exhibit a non-trivial relative risk separation. Building on these findings, we propose Discordance-Oriented Causal Subgraph Discrimination (DCS), which uses full-graph classification risk as an observable proxy for latent discordance strength and adaptively enforces a stronger predictive advantage of candidate causal subgraphs over their non-causal complements, enabling efficient and robust causal subgraph identification. Extensive experiments on 8 datasets demonstrate consistent improvements over existing graph OOD methods.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
Graph neural networks, Out-of-distribution

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WzrrOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In out-of-distribution settings, existing GNN methods typically exploit the extraction of causal subgraphs in order to improve generalization capabilities which ignores the cross-sample generalization cues provided by the spurious subgraphs (complement of the causal subgraph). In this paper, the authors focus on concept shift and find that spurious subgraphs can provide predictive cues that conflict with the predictive cues provided by the causal subgraphs (termed cue discordance). The authors find that when this phenomenon occurs, predictive accuracy suffers. Specifically, the ratio of discordant samples set scales roughly linearly with OOD accuracy. The authors then prove that, on discordant samples, a classifier that only sees the causal subgraph must have strictly lower prediction loss than a classifier that only sees the spurious subgraph by some margin (D3C criterion). Based on this theory, the authors propose the DCS method which includes a learnable edge mask that splits each input graph into a candidate causal subgraph and a candidate spurious subgraph and a shared predictor that outputs predictions for the candidate causal subgraph, the candidate spurious subgraph, and the entire input graph. Then, the model is trained via a three component loss function: CE on candidate causal subgraph and full graph, hinge term that is high when causal subgraph CE isn't lower than spurious subgraph CE, and a compactness regularization term. The final prediction is done on the extracted candidate subgraph. The method is evaluated on the GOOD benchmark (concept shift) and achieves SOTA results.

Critical strengths
  1. The cue discordance analysis on OOD generalization for graphs is novel and interesting. It provides a fresh viewpoint on why models misclassify OOD samples and how causal subgraphs should be extracted. It is also supported by strong empirical findings that shows predictive accuracy scales roughly linearly with the ratio of discordant samples.
Critical weaknesses
  1. The existence of discordant samples in the training set is not supported. The paper provides evidence that there are significant amounts of discordant samples in the test set of Motif-basis and CMNIST-color (Figure 2). However, their theory and proposed method depends on the existence and prevalence of discordant samples in the training set. Specifically, if there are very few (or even zero) discordant samples in the training set, their novel relative-risk loss term will be almost constant across all training samples. Then, the training loss would degenerate to only the task loss term and the compactness regularization term. For example, in the default Motif-basis dataset, ~97% of the training samples are forced to have a spurious subgraph that is correlated with the label which means there are at most 3% (practically ~2%) discordant samples in the training set. In this case, the relative-risk term of the proposed loss function (second term) would contribute very little to model training.
  2. There are inconsistencies with the reported results in the main results table (Table 1). The reported results for some of the baselines are inconsistent with their original papers. Most notably, for PrunE, the original paper reports 79.50±1.57 and 90.28±1.72 on HIV-size and Motif-basis (Yao et al, 2025) which would both be significantly higher than the proposed method; while this paper reports 69.73(0.79) and 77.63(1.74) for PrunE respectively. I haven't checked every baseline but IRM, ERM, and iMoLD results also seem to be somewhat inconsistent with their original papers (Gui et al, 2022; Zhuang et al, 2023).

Minor:

  • The CMNIST-color dataset contains 10 digits and 10 colors (Gui et al, 2022) which means the test set should have ~90% discordant samples yet Figure 2 reports ~67%.
Questions for authors

The questions correspond to the weaknesses.

  1. Could the authors provide evidence that demonstrates the existence/prevalence of discordant samples in the training set of the real datasets? Furthermore, could the authors provide some additional analysis on the scale and contribution of each loss term?
  2. Could the authors account for the discrepancies between the baseline results in Table 1 and the results reported in the corresponding original papers?
7 comments

We thank for your valuable comments.

  • W1. Discordance Discrimination

We thank the reviewer for raising this concern. We clarify that our method is motivated by the generalizability of non-causal subgraphs, rather than being designed for a particular distribution of discordant training samples. Since non-causal subgraphs can generalize across samples, they may induce classifiers to rely on spurious predictive cues instead of causal mechanisms, even when these cues largely agree with the ground-truth labels during training.

Importantly, cue discordance can persist even under strong spurious correlations, as explaining the labels perfectly is impossible for non-causal cues. Although such conflicts may be rare during training, distribution shifts can substantially amplify them at test time, exposing the limitations of spurious reliance. Cue discordance thus reveals the competition between causal and non-causal predictive mechanisms, rather than serving as a prerequisite for our optimization.

Specifically, Eq.17 contrasts the predictive losses of selected and complementary subgraphs on all training samples, with discordance-aware weighting only adjusting their relative importance. Even if the discordance-aware weights become uniform, the objective reduces to an unweighted relative-risk contrast rather than constant.

Motif-basisMotif-sizeCMNIST-colorHIV-size
w/o weighting80.6467.5349.3770.06
w/o margin81.2668.9348.8271.54
Original84.8672.0854.1374.47

Together with the ablation above, these results confirm relative-risk discrimination provides meaningful optimization signals beyond discordant samples, enabling the model to address competing spurious cues during training before their adverse effects become more pronounced under distribution shifts.

  • W2&Q2. Reproducibility

For ERM and IRM, we used the official GOOD implementations, obtaining results consistent with those originally reported in the GraphOOD benchmark. For PrunE and iMoLD, we adopted the official implementations and settings provided by thepaper. Furthermore, we conducted a more extensive hyperparameter search for our method, achieving better performance than their originally reported results:

Motif-basisMotif-sizeHIV-scaHIV-size
reported ERM81.3170.8372.4861.91
reported IRM80.3770.1571.81-
reported PrunE90.28--79.5
reported iMoLD--74.3277.43
Ours92.1472.0876.5481.33

Moreover, minor discrepancies may arise from randomness in data splitting on real-world datasets, as the concept-shift setting involves stochastic partitioning based on domain–label correlations. For synthetic datasets, the random graph generation process may also introduce additional variability beyond data splitting.

  • Q1. Discordant Evidence and Optimization

Importantly, spurious correlations in real-world datasets are statistical rather than deterministic. For example, in the SST2-length training set, negative samples are substantially longer than positive ones on average (15.74 vs. 5.05). Nevertheless, 25.53% of negative samples fall below the overall average length, while 7.89% of positive samples exceed it. This demonstrates that spurious length cues do not consistently align with ground-truth labels, providing evidence of cue discordance within the training distribution:

LengthProportion#SamplesNegativePositive
≤542.84%11,6837.67%92.33%
6–1536.36%9,91662.10%37.90%
>1520.80%5,67188.54%11.46%

Moreover, real-world cue discordance can arise from multiple interacting spurious factors and cannot be fully characterized by a binary partition based on a single attribute. This further motivates our soft weighting strategy, which uses full-graph classification loss as an observable proxy to adaptively emphasize samples with potentially stronger cue conflicts, rather than requiring explicit identification of discordant samples.

Regarding the contribution and scale of each loss term, we present the ablation results and the loss magnitudes at different training stages on Motif-basis. These analyses help clarify how the different objectives jointly influence optimization beyond their effects on final predictive performance:

VariantMotif-basisMotif-sizeCMNIST-bgHIV-size
w/o compactness loss81.63(-3.23)69.54(-2.54)50.51(-3.62)72.39(-2.08)
w/o weighting80.64(-4.22)67.53(-4.55)49.37(-4.76)70.06(-4.41)
w/o relative contrast79.27(-5.59)64.75(-7.33)47.87(-6.26)69.59(-4.88)
w/o margin81.26(-3.60)68.93(-3.15)48.82(-5.31)71.54(-2.93)
w/o whole discrimination loss74.23(-10.63)60.57(-11.51)43.29(-10.84)69.32(-5.15)
Original84.8672.0854.1374.47
Training Epoch50100150200
Weighted Relative Loss0.16940.1490.13720.1347
Compactness Loss0.00540.00240.00130.0012
Classification Loss0.39930.32760.31270.3013

W1/Q1. Thank you for clarifying the motivation of the method and how the relative-risk term acts on all training samples. I consider this point addressed.

W2/Q2. The original discrepancies between the reported results and Table 1 of the paper (PrunE: 90.28 -> 77.63 on Motif-basis and 79.50 -> 69.73 on HIV-size, iMoLD: 77.43 -> 67.87 on HIV-size) are still not explained. Randomness in data splitting or graph generation does not explain gaps of this size, given that your reported standard deviations are 1 to 2 points. Could the authors explain what accounts for these differences?

Furthermore, Sec 5.4 claims the method is stable over a broad range of hyperparameter values (λ_rel and δ), yet the rebuttal reports significant increases from 84.86 to 92.14, 72.56 to 76.54, and 74.47 to 81.33 on Motif-basis, HIV-scaffold and HIV-size from a more extensive hyperparameter search. Was the new search over λ_rel and δ? If so, how is this consistent with the stability claim? If not, then which hyperparameters did it cover? The GOOD benchmark fixes the general training settings (e.g., layer count and hidden dimension) and tunes only method-specific hyperparameters. If the new search departs from that without re-running the baselines under the same protocol, the comparison is not fair. Could the authors provide the exact tuning protocol for the proposed method and the evaluated baselines?

We thank the reviewer for raising these important concerns. We carefully provide two major sources of discrepancies between our reported results and those in the original papers.

(1) Motif dataset: Differences in dataset generation.

PrunE uses the pre-generated Motif dataset downloaded from the official GOOD repository, whereas our original experiments regenerated the dataset using the original GOOD Benchmark code. We verified that our generation code and configurations were consistent with those provided by GOOD. Importantly, all results reported in our original paper were obtained using this same regenerated dataset, ensuring consistent experimental conditions across methods. In our additional experiments, we adopted the officially released GOOD dataset, obtaining an accuracy of 92.14% on Motif-basis. We also observed improvements for other baselines, including ERM (78.66 → 81.31 and 67.77 → 70.83) and IRM (78.67 → 80.37 and 67.11 → 70.15).

(2) HIV dataset: Differences in molecular feature processing.

We further identified important differences between the original implementations of PrunE/iMoLD and the standard GOOD Benchmark pipeline:

  • Molecular encoders: Both PrunE (gcn.py) and iMoLD (models/gnnconv.py) employ different specialized atom encoders that are not included in the standard GOOD implementation. These encoders explicitly process molecular atom and bond attributes, whereas our original implementation (for all methods) directly projected the input features into a 300-dimensional hidden space without such specialized molecular encoding.
  • Additional handcrafted features: iMoLD further extracts 38 handcrafted molecular features, introducing additional feature engineering beyond the standard GOOD pipeline.

Since these preprocessing and encoding strategies are not part of the standard GOOD implementation and differ across methods, our original experiments adopted a unified feature-processing pipeline rather than incorporating these method-specific modifications, with the aim of maintaining consistent input processing across methods.

We appreciate the reviewer for highlighting the importance of explicitly distinguishing benchmark-standardized comparisons from faithful reproductions of the original methods.

Reviewer_42FrOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how graph neural networks can reduce reliance on spurious correlational structures under distribution shifts.The authors focus on "cue conflict": the causal subgraph within a graph supports the correct class, while the spurious subgraph favors an alternative class. Controlled experiments reveal that prediction accuracy degrades as the severity of conflict increases. Under certain assumptions, the authors further prove that causal subgraphs enjoy a relative predictive risk advantage in conflict settings.Accordingly, they propose DCS. The method decomposes a graph into a candidate causal subgraph and its complementary component. A shared predictor is used to compare predictions from these two parts. A margin loss is applied to enforce superior performance of the candidate causal subgraph. Meanwhile, the full-graph loss upweights training samples suspected of suffering from cue conflict.

Critical strengths

1.It presents a clear problem formulation and analyzes cue conflict via controlled experiments. By varying the proportion of conflicting samples on the Motif and CMNIST datasets, the paper demonstrates its correlation with declining prediction accuracy and rising classification loss. It further investigates prediction errors on conflicting samples, furnishing empirical evidence for method design.

2.The theoretical analysis is clearly connected to the method design. Under certain assumptions, the paper analyzes the expected prediction losses of causal and non-causal subgraphs and uses this analysis to design a relative loss that guides subgraph extraction. The main experiments and ablation studies also support the effectiveness of this design.

Critical weaknesses

1.The authors use the full-graph classification loss as a soft proxy for cue discordance and acknowledge that it may also reflect general sample difficulty. However, high loss may arise from factors unrelated to discordance, such as structural complexity or difficulty in recognition. Although controlled experiments show that the average classification loss increases as the proportion of discordant samples rises, the current theory and experiments do not sufficiently establish the extent to which the performance gains from weighting result from mitigating cue discordance rather than learning from generally difficult samples.

2.The comparison with directly related methods remains insufficient. For example, existing methods such as VIVACE [1] and DisC [2] already combine subgraph decomposition with loss weighting, but this paper has not sufficiently discussed the connections and differences between these methods and the approach used in DCS. Furthermore, the related work section mainly describes existing approaches without explaining their limitations in detail, and a targeted analysis of how the design of DCS alleviates these shortcomings is still lacking. [1]Li H, Wang X, Zhu X, et al. Disentangling invariant subgraph via variance contrastive estimation under distribution shifts[C]//Forty-second International Conference on Machine Learning. 2025. [2]Fan S, Wang X, Mo Y, et al. Debiasing graph neural networks via learning disentangled causal substructure[J]. Advances in Neural Information Processing Systems, 2022, 35: 24934-24946.

Questions for authors

1.Can the existing theory or evidence show that the performance gains from weighting based on the full-graph loss stem from mitigating cue discordance rather than placing greater emphasis on learning from generally difficult samples?

2.Compared with methods that also combine subgraph decomposition with loss weighting, such as VIVACE and DisC, what are the key new mechanisms introduced by DCS? Which specific limitations of existing methods do these mechanisms alleviate, and what theoretical or empirical evidence supports this?

3 comments

We thank for your valuable comments.

  • W1&Q1. Discordance Weighting

To mitigate the concern that high classification loss may arise from factors unrelated to cue discordance, DCS employ a warm-up before activating the relative objective, allowing the model to establish reliable predictive representations and reducing the influence of uninformative high losses during early training. Furthermore, weight clipping and stop-gradient improve optimization stability.

Importantly, extended experiment results on misclassifications also show that DCS reduces the discordance-induced error gap from 23.81 pp to 13.28 pp, providing empirical evidence that its improvements are closely associated with mitigating cue discordance, supporting its practical effectiveness of mitigating cue discordance rather than learning from generally difficult samples (lower is better):

MisclassificationsNon-disc. (%)Disc. (%)Gap (pp)Non-disc. (#)Disc. (#)Gap
ERM6.5530.3623.8113212101078
PrunE6.0128.6322.6212111411020
Ours6.3119.5913.28127781654

We also clarify that weighting alone is not intended to mitigate cue discordance. Instead, full-graph loss serves as a soft proxy for identifying potentially discordant samples, while the relative margin objective explicitly encourages causal–spurious discrimination. These two components work together: weighting determines where to emphasize, whereas relative discrimination determines what to enforce. This is also supported by the ablation results:

Motif-basisMotif-sizeCMNIST-colorHIV-size
w/o weighting80.6467.5349.3770.06
w/o margin81.2668.9348.8271.54
Original84.8672.0854.1374.47
  • W2&Q2. Compared to Related Works

We add comparisons with DisC and VIVACE. The results are presented below, demonstrating that our method substantially outperforms both approaches:

Motif-basisMotif-sizeCMNIST-colorCMNIST-bgHIV-scaHIV-sizeSST2Twitter
DisC51.0850.3924.9924.3168.0758.7676.5148.61
VIVACE76.4767.3537.0439.8769.5766.3477.6351.74
Ours84.8672.0854.1354.3472.5674.4778.7156.67

The key distinction lies in how non-causal subgraphs are characterized and utilized. Unlike DisC and VIVACE, DCS explicitly exploits cue discordance arising from the nontrivial generalizability of non-causal subgraphs, where elevated full-graph loss can signal competing causal and spurious predictive cues.

DCS addresses this through two mechanisms:

  1. Discordance-conditioned relative-risk discrimination. Unlike DisC, which relies on a bias-focused predictor for reweighting, DCS explicitly enforces a predictive-risk advantage of the selected subgraph over its non-causal complement, supported by Theorem 1 and Proposition 1. Moreover, while DisC relies on restrictive assumptions regarding the relationship between (H(C|Y)) and (H(S|Y)), our analysis requires no predefined ordering between them.
  2. Discordance-aware optimization. Unlike VIVACE, which relies on variant-subgraph augmentation and inverse propensity weighting, DCS directly uses full-graph loss as a soft signal to emphasize likely discordant samples, without additional spurious-predictor training or propensity estimation, aligning the training objective more closely with the desired subgraph discrimination. Furthermore, VIVACE does not explicitly establish the statistical fidelity of its estimated propensity scores, and its weighting formulation may emphasize samples with lower spurious-predictor loss, potentially conflicting with its debiasing objective.

I appreciate the authors’ responses. However, after careful consideration, I have decided to maintain my original score, as I still believe that my original overall score best reflects the quality of this paper.

Thank you for your thoughtful consideration of our responses. We respect your decision and appreciate your feedback. May we kindly ask whether there are any further concerns that our rebuttal has not adequately addressed? We would be glad to provide further clarification.

Reviewer_4MdpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies cue discordance in graph OOD generalization, defined as samples in which the causal subgraph and the spurious subgraph tend to support different labels. Under the conditions of causal sufficiency, Bayes decision consistency, and the multiclass Tsybakov margin condition, this paper derives the conditional cross-entropy risk gap between the causal subgraph and the non-causal subgraph on inconsistent samples. It proposes the DCS method, which learns a Top-r edge mask and enforces a margin between the selected-subgraph loss and the complementary-subgraph loss, where this margin is weighted by the detached full-graph loss.

Critical strengths

S1. The proposed method uses a relatively simple training objective and does not require environment annotations or complex data augmentation.

S2. The results are promising.

S3. This paper is easy to follow.

Critical weaknesses

W1. The theoretical results do not establish the causal identification interpretation motivating Eq. (15). Theorem 1 concerns the true causal decomposition under a Bayes-decision-consistent causal predictor, while Proposition 1 provides a surrogate upper bound; neither establishes that optimizing the criterion over candidate subgraphs recovers causal structure or excludes spurious solutions.

W2. As written, Eq. (17) can be reduced by increasing the complementary-view loss through the shared predictor, rather than improving the extracted subgraph.

W3. Figure 2 establishes an aggregate association between discordance prevalence and loss, but does not directly validate full-graph loss as a sample-level weighting signal during training.

Questions for authors

Q1. What objective is used in the “w/o RC” ablation?

Q2. Does the complementary-view loss update the shared predictor through the relative objective?

2 comments

We thank for your valuable comments.

  • W1. Theoretical Scope

We clarify that our theoretical results characterize the relative-risk advantage of causal subgraphs under cue discordance, providing a principled motivation for the discrimination objective in Eq. (15). Importantly, our practical optimization does not rely on the relative-risk criterion alone. We first employ a supervised warm-up to establish meaningful predictive representations, making the sample-wise loss comparisons in Proposition 1 more informative for subsequent relative discrimination. Subsequently, task supervision and relative discrimination jointly encourage the selected subgraph to retain label-predictive information while outperforming its complement.

Empirically, our ground-truth subgraph evaluation demonstrates that our method more effectively recovers causal subgraphs than competing methods, achieving the highest edge-level scores:

MethodEdge Acc.Edge RecallEdge Prec.Edge F1
GIL0.8920.400.720.51
GSAT0.9040.440.740.55
LECI0.9160.490.760.60
Ours0.9340.570.790.66

Regarding excluding spurious solutions, we further report the causal subgraph recovery performance (F1 score) under varying strengths of spurious correlations. The results demonstrate that our method consistently achieves the most accurate recovery of causal subgraphs compared to the baseline (without the discrimination objective), indicating its superior ability to mitigate the influence of spurious solutions:

Spurious Rate(Motif-basis)0.60.70.80.90.95
Baseline0.6430.6390.6310.6240.607
Ours0.7260.7170.7090.6830.664

Moreover, under Theorem 1, the real causal decomposition satisfies the desired relative-risk ordering under cue discordance, whereas selecting the spurious component while retaining the causal component in its complement violates this ordering and incurs a larger relative-risk penalty. Thus, the proposed criterion is intended to discourage such spurious solutions by encouraging the desired relative-risk ordering.

  • W2. Loss Design

To directly examine the concern about Eq.17, we further prevent the complementary-view loss from updating the shared predictor, while preserving its gradients to the extractor. The resulting performance remains comparable to the original DCS, indicating that the improvements do not rely on directly degrading complementary-view predictions through predictor updates:

Motif-basisMotif-sizeCMNIST-colorCMNIST-bgHIV-scaHIV-sizeSST2Twitter
Original84.8672.0854.1354.3472.5674.4778.7156.67
Frozen84.4372.5753.6253.7572.9374.3478.0356.42

Moreover, DCS optimizes Eq.17 jointly with supervised losses on both the extracted subgraph and the full graph through a shared predictor. These constraints discourage arbitrary predictor degradation. Importantly, an increased complementary-view loss does not necessarily indicate predictor degeneration; it can also reflect the extractor successfully concentrating label-relevant structures in the selected subgraph. Furthermore, the causal subgraph evaluation also demonstrates improved causal subgraph identification, providing empirical support beyond predictive performance.

  • W3. Weighting Signal

We first clarify that our ablation results provide direct empirical evidence for the effectiveness of sample-level weighting, consistently improving performance across all datasets:

Motif-basisMotif-sizeCMNIST-colorHIV-size
w/o weighting80.6467.5349.3770.06
Original84.8672.0854.1374.47

Furthermore, the controlled setting in Figure 2 provides meaningful evidence for sample-level weighting. Since cue discordance is binary at the sample level, its relationship with classification loss is examined statistically by varying its prevalence across the dataset. In Figure 2, the discordance ratio is the only variable being deliberately manipulated, while the sample distributions remain fixed. The increasing average full-graph loss indicates higher expected losses for discordant samples, supporting its use as a soft weighting signal. Importantly, our method does not require identifying individual discordant samples, but simply assigns higher weights to samples more likely to be discordant, as clarified in Appendix A.

  • Q1.w/o RC

'w/o RC' means removing the complementary-view contrast term while retaining discordance-aware weighting, yielding Lrel=1B∑iwiℓic\mathcal{L}_{rel}=\frac{1}{B}\sum_i w_i\ell_i^c.

  • Q2.Optimization

Yes, the relative objective updates both the extractor and the shared predictor through the loss. Moreover, as shown in W2, blocking predictor gradients from the complementary-view loss still yields comparable performance.

Dear Reviewer,

Thank you again for your valuable feedback. We would like to kindly ask whether our rebuttal has adequately addressed your concerns, or if there are any remaining questions that we could further clarify.

We would be happy to provide additional explanations if needed, and we sincerely appreciate your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.