acceptodds
Under review as a conference paper at ICLR 2027

Graph Unsupervised Domain Generalization via Environment Confusion

Abstract

Graph-level out-of-distribution (OOD) generalization is critical for deploying graph neural networks in real-world applications where distribution shifts are inevitable. Existing approaches typically require class labels during representation learning to decide which substructures are predictive, limiting their applicability. We propose **SCARCE**, a label-free representation-learning framework that yields distribution-shift-robust graph representations from an environment partition of the training graphs alone, whether that partition comes from dataset metadata or, when none is available, from clustering a self-supervised embedding. Our method operates in two stages: (1) a variational graph autoencoder is warm-started for graph reconstruction; and (2) an adversarial confusion mechanism forces the encoder to produce representations from which a discriminator cannot distinguish environments, while preserving graph-reconstruction fidelity. We provide theoretical analysis showing that environment confusion contracts environment-discriminative directions in the representation space, which, under regularized downstream classifiers, is equivalent to imposing a stronger directional penalty on environment-associated features. This feature-reweighting mechanism reduces reliance on spurious correlations without erasing information, offering a label-free, continuous relaxation of last-layer reweighting methods. Experiments on a controlled synthetic model, the GOOD-Motif benchmark, and molecular property prediction under assay, scaffold, and size shifts show that **SCARCE** achieves state-of-the-art OOD generalization among label-free methods, using no class labels during representation learning.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
graph representation learning, out-of-distribution generalization, self-supervised learning

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_7xhfOpenReview ↗
Rating
2(+1)
Confidence
4
Informativeness
2
Presentation
2(−1)
Soundness
2(+1)
Read review
Paper summary

The authors propose SCARCE. A framework which operates in two cycles to improve OOD generalization for a VGAE encoder. The first, involves a self-supervised contrastive encoder applying reconstruction loss to separate the invariant representation from any shortcuts in the environment, making the representation fixable before stage two. The second, applies pseudo-environments from the frozen anchor clustering into an adversary which predicts the environment as optimized by a uniform confusion loss that can improve alignment with invariants and reduce spurious correlation to the environment.

Critical strengths

The most significant strength of SCARCE is it's capacity to improve downstream OOD performance for the realistic DrugOOD IC50 and EC50 Assay, Scaffold, and Size shifts under the author's proposed learning objective. I believe that this reflects a clever formulation for label-free learning objectives without accidentally exposing testing leakage. I also appreciate the author's efforts to establish clarity within their writing.

Critical weaknesses

I have several concerns, the largest regarding the soundness of the label-free objective, as described in Q1. Largely, I find the consistency of the label-free objective proposed within Equation 2 to be consistent with the standard labelled invariant learning objectives. However, the author propose SCARCE to function in an unlabelled invariant learning environment which is proposed as a means for not testing standard invariance learning methods such as IRM. This is largely concerning since it makes it difficult to consider that a label-free unsupervised OOD problem is sound, especially given that the DrugOOD dataset provides labels for supervised learning. Additionally, the title for the submitted PDF is different than the title listed within OpenReview.

Questions for authors
  1. This concern is based on my description in the 'Critical Weaknesses' section and the exactness of the statement in Line 47: "Class labels are the expensive part". Couldn't existing invariant learning methods, like IRM, still work with pseudo-environment labels from MVGRL clustering? Or is it unreasonable to treat the pseudo-environments as potential Y label for IRM or other invariance learning techniques? I ask largely because Assay, Scaffold, and Size shift for EC50 and IC50 are dependent on molecular structure or number of nodes, which when encoded could serve a kNN-derived labels. So, it follows that the label-free learning objective within this study (Equation 2) losses it soundness as a practical OOD objective given the potential viability for kNN-derived labels as a target for IRM training.
  2. Is Line 5 within Algorithm 1 also Equation 7 or is the 'Compute' function dependent on MVGRL clustering as Equation 4, Lines 158, 454-455, and 478 lead me to assume? This affects my current interpretation of how SCARCE determines the proper separation between S and C since it seems that SCARCE relies on MVGRL without explicitly stating it in the current manuscript.
  3. How does the controlled contraction for geometry overlap and probe regularization function within the shifts of EC50 and IC50? I understand that both datasets are unlikely to indicate as clean separation within Figure 3, but it could be an interesting diagnostic to determine whether the larger success of the Scaffold and Assay split tests is correlated with cleaner separation for S and C. This would fully convince me that the theory proposed within this work is sound.
1 comment
Reviewer_7xhfOpenReview ↗

Rebuttal Response

Thank you to the authors for their rebuttal, I appreciate the clarification on my second and third questions. I have raised my score to reflect that those concerns are addressed on those specific points. I am pleasantly surprised at the effectiveness of SCARCE to manage the proposed label-free objective and believe that the additional rebuttal experiments can be interpreted for useful insight in applying environment confusion to OOD generalization.

On my first concern, I clarify that I do not believe that the entire label-free setting is unsound. My assertion is grounded in how the presentation of the proposed evaluation objective, as it currently stands, in-context to invariant learning and self-supervised/unsupervised settings is confusing. Equation 2 describes a scenario including YY for minimizing risk between environments, just like invariant learning, but then immediately claims that the objective is label-free. I understand this could be done for multiple reasons, including: for effect, is related to experimental setup (Line 361), and that the remainder of Line 180 follows self-supervised learning paradigms. But, it currently reads strangely to only indirectly mention self-supervised learning within Lines 175-184, especially given that SCARCE is centered around a self-supervised/unsupervised learning paradigm.

From a presentation standpoint, a simple one-line update within Lines 175-184 which is explicit about the self-supervised/unsupervised nature of this work would clarify the intent of Line 180 and make the ruling out of existing invariant learning methods no longer read as awkward to me.

From a practical standpoint, my concern was also largely-grounded in that this work's tested IC50 and EC50 DrugOOD datasets do not inherently suffer from a lack of labels, where previous OOD-generalization techniques with full label-access have demonstrated strong performance [1,2]. It is not necessary for additional benchmarking experiments on the two datasets with CIGA and GALA, but clarification on the benefits that SCARCE provides over models like CIGA and GALA will be of great benefit to demonstrating the value of SCARCE's contribution.

I appreciate the author's consideration and look forward to continued discussion.

[1] Chen, Y., Zhang, Y., Bian, Y., Yang, H., Kaili, M.A., Xie, B., Liu, T., Han, B. and Cheng, J., 2022. Learning causally invariant representations for out-of-distribution generalization on graphs. Advances in Neural Information Processing Systems, 35, pp.22131-22148.

[2] Chen, Y., Bian, Y., Zhou, K., Xie, B., Han, B. and Cheng, J., 2023. Does invariant graph learning via environment augmentation learn invariance?. Advances in Neural Information Processing Systems, 36, pp.71486-71519.

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

This paper proposes a label-free framework for graph out-of-distribution generalization. The method first employs a variational graph autoencoder to retain both invariant and environment-related information through graph reconstruction, and then introduces adversarial environment confusion to suppress environment-discriminative directions in the latent space. Theoretical analysis establishes a connection between representation contraction and directional regularization in downstream linear classifiers. Experiments on synthetic data, GOOD-Motif, and DrugOOD demonstrate the effectiveness of the proposed approach.

Critical strengths
  1. The paper addresses a meaningful problem by reducing the reliance on class labels for graph OOD representation learning, which is particularly valuable in scenarios where labeled data are scarce or expensive to obtain.

  2. The theoretical analysis provides useful insights by establishing a connection between representation contraction and anisotropic regularization, offering an interpretation of how suppressing environment-discriminative directions can improve OOD generalization.

Critical weaknesses
  1. The proposed pseudo-environments are obtained by clustering self-supervised graph representations, but there is no guarantee that these clusters capture the real environmental factors responsible for OOD shifts rather than causal invariant, task-relevant information. This raises concerns about the reliability of the subsequent environment confusion mechanism.

  2. The DFR experiment in Figure 2 constructs a group-balanced subset using both class and environment labels from the validation set. However, the main SCARCE evaluation fits linear probes using labeled training data without such group balancing. This discrepancy raises questions about whether the DFR results adequately support the claimed retention of invariant features under the actual evaluation setting.

  3. While Theorem 4.1 establishes information retention under reconstruction, it does not guarantee that the invariant factor (C) remains preserved after adversarial environment confusion. (C) and the spurious factor (S) may be highly entangled in the latent space, suppressing environment-discriminative directions could also remove task-relevant information. The theoretical analysis does not sufficiently address this possibility.

  4. Figure 1 is overly cluttered, with inconsistent visual elements, many of which appear rasterized and become blurry when enlarged. A cleaner, vector-based illustration would improve readability and presentation quality.

Questions for authors
  1. The proposed framework appears largely similar to conventional domain-adversarial representation learning, combining reconstruction with adversarial environment confusion. What fundamentally distinguishes SCARCE from existing approaches, particularly in its ability to suppress spurious shortcuts rather than merely reduce environment predictability?

  2. Most baseline methods considered in the experiments were proposed in or before 2024. Could the authors include more recent and competitive approaches to better establish the effectiveness of SCARCE?

Reviewer_5BsdOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes SCARCE, a label-free framework for graph OOD representation learning. It combines graph reconstruction with adversarial environment confusion and supports both supplied and inferred environment partitions. The authors analyze information retention and representation contraction, and connect contraction to regularized linear classification. Experiments on synthetic data, GOOD-Motif, and DrugOOD show competitive performance, particularly on molecular distribution shifts.

Critical strengths
  1. The paper addresses graph OOD representation learning without class labels in the representation-learning objective. Combining reconstruction with environment confusion accommodates both supplied and inferred environment partitions, broadening the applicability of the approach when task annotations are limited.
  2. SCARCE achieves strong results across most DrugOOD settings, particularly under scaffold shifts. Its performance with limited labels supports the practical value of the approach.
Critical weaknesses
  1. Appendix A.2 analyzes a fixed linear adversary with direct representation updates, so its connection to the implemented GNN training remains approximate. The experiments provide limited direct evidence of selective contraction and task-information retention after confusion, especially with clustered pseudo-environments. This weakens the mechanistic explanation of the observed gains.
  2. Although the paper describes validation-based model selection, the DrugOOD linear probe's regularization strength and selection procedure remain unclear. Since regularization is central to the proposed mechanism, these missing details limit reproducibility and assessment of the reported gains.
Questions for authors
  1. Can the authors provide direct evidence that environment confusion contracts spurious directions while preserving task information in the trained GNN, particularly when pseudo-environments are obtained by clustering?
  2. Could the authors clarify the difference between the IC50-scaffold results in Table 2 and the corresponding results in Table 4? Are the evaluation protocol and aggregation procedure the same?
  3. What regularization strength and selection procedure are used for the DrugOOD linear probes? For GOOD-Motif, which data split and metric determine the best epoch reported in Table 6?
1 comment

Thank you for the thoughtful responses and additional experiments. The clarifications on how Tables 2 and 4 aggregate results, how the DrugOOD probes are regularized and selected, and how the GOOD-Motif checkpoint is chosen address my questions in Q2 and Q3.

On Q1, the EC50 results show that pseudo-environment membership explains a smaller share of the variance after confusion training and remains linearly decodable. The clustered partition also yields higher OOD AUC than the random partition on all three EC50 splits. These findings strengthen the evidence for contraction and for the value of clustering. As you note in Part 2, however, a clustered partition can capture both spurious variation and stable task information. The variance measure tracks that partition as a whole, and the AUC measures their net effect on prediction. This leaves the relative attenuation of the two only indirectly assessed.

The matched-ensemble results also clarify the scope of the predictive gains: SCARCE is close to the strongest reported baseline on IC50-scaffold and falls below it on IC50-size. I appreciate the additional evidence. Given the importance of selective contraction to the paper’s explanation of its DrugOOD results, I will maintain my current rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.