acceptodds
Under review as a conference paper at ICLR 2027

ITXGNN: Learning Stable Explanations in Self-Interpretable Graph Neural Networks

Abstract

Self-Interpretable Graph Neural Networks have become quite effective at being able to retain the prediction accuracy of comparable black box methods while also introducing interpretability to their decision making processes. However, most self-interpretable methods do not generate stable explanations. In this work, we define a key notion of stability that is absent in the literature and provide a framework that addresses this problem head on. Our framework Information-Theoretic Self-Explainable Graph Neural Network (ITXGNN) is a method that utilizes subgraph extraction with prototypes, clustering, and representation learning in order to satisfy an objective function that prioritizes stability while maintaining accuracy. In addition to our newfound definition of stability we provide theoretical justification of our objective function and derivations highlighting the natural choice of objective function as well as a mathematical bridge from theory to implementation. We conduct comprehensive experiments on several datasets showcasing our method's strength on stability in comparison to leading existing self-interpretable GNNs while maintaining competitive prediction performance. We also conduct ablation studies that break down our frameworks key components and provides quantitative analysis for each components contribution and we also provide a qualitative case study showcasing explanation quality.

Primary area:
interpretability and explainable AI
Keywords:
Self-Interpretable Graph Neural Networks, Graph Explainability, Explanation Stability, Prototype Learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_8qHmOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes ITXGNN, a self-interpretable GNN designed to improve explanation stability across graph instances. It defines stability through intra-group consistency and inter-group discrimination, and combines subgraph extraction, learnable clustering, and prototype-based representation learning with an information-theoretic objective. Experiments demonstrate improved stability metrics while maintaining competitive predictive performance.

Critical strengths

S1. The paper addresses an important problem by explicitly considering both intra-class explanation consistency and inter-class discrimination. S2. ITXGNN achieves substantial improvements in explanation stability across synthetic and molecular graph benchmarks. S3. The paper provides theoretical analysis, ablation studies, and qualitative examples to support its methodology.

Critical weaknesses

W1. Definition 1 uses expected pairwise explanation distances, whereas the molecular experiments compute FID between randomly divided subsets, which does not directly measure within-class explanation variability. W2. Maximizing mutual information between prototypes and clusters does not necessarily guarantee structural consistency among extracted subgraphs, and Theorem 1 only establishes a bound in the embedding space. W3. The experiments lack systematic evaluation of explanation faithfulness and ground-truth overlap, making it difficult to determine whether more stable explanations are also more accurate. W4. The ablation study is limited to BA-2Motif, and the improvement over frozen clustering is small relative to the reported standard deviations.

Questions for authors

Q1. In Section 6.1, why is FID between two random subsets used to approximate the intra-group dispersion in Definition 1? Since both subsets come from the same distribution, their FID can approach zero even when individual explanations differ substantially. Could the authors report direct pairwise explanation distances? Q2. How does the objective in Equation 3 ensure that extracted subgraphs are structurally stable rather than merely producing well-separated embedding clusters? Can the authors empirically verify the relationship between prototype separation and subgraph-level stability? Q3. Could the authors report explanation F1/Jaccard against ground-truth motifs and faithfulness metrics under matched explanation-size budgets? This would clarify whether the stability improvements correspond to identifying more relevant subgraphs. Q4. How are the cluster assignment probabilities (q_{ik}) learned and inferred for unseen graphs? Are prototypes recomputed from batch embeddings or directly optimized using the prototype learning rate reported in Appendix D? Q5. Could the authors extend Table 4 to molecular datasets and compare against simpler class-conditioned contrastive or prototype-based regularization using the same backbone?

2 comments
AuthorsOpenReview ↗

Response to Reviewer 8qHm (1/2)

We would like to thank the reviewer for taking the time and effort to read and understand our work, we appreciate the constructive feedback.

Q1. We agree, and we now report direct pairwise explanation distances. For all methods on all four molecular datasets, we compute the average within-class and between-class pairwise distances in ECFP and GIN embeddings and report their ratio (Table R1). ITXGNN has the lowest ratio on every dataset in both spaces: 0.43–0.84, vs 0.70–1.05 for the baselines. Its same-class explanations are therefore closer to one another than to other-class explanations under the quantity Definition 1 defines. FID is kept as a supplementary distribution-level analysis which we will move to the appendix.

Q2. The objective acts on embeddings and does not by itself guarantee structural stability. Embedding stability is nevertheless the right target, for two reasons. First, structural stability cannot be optimized directly: isomorphism and edit distance are discrete and non-differentiable, whereas agreement in embedding space can be trained end to end. Second, the embedding is what the model predicts from so explanations that agree in embedding space mean that same-class graphs are classified on the same evidence. Because the extractor is trained through the same objective, the simplest way to make a cluster's embeddings agree is to select the shared structure; mapping structurally different subgraphs to the same point would require the encoder to discard their differences. Empirically, this is what happens. On BA-2Motif, the IsoRate of 0.970 shows that embedding clusters coincide with isomorphic subgraphs. To verify the relationship directly, we have conducted a new experiment; where we perturb the learned explanations of a trained model by adding or removing nodes or replacing them with a different subgraph (Table R7). The distance to the class centroid grows with the structural deviation, from 0 for the learned explanation to 0.066 after adding five nodes, 0.210 after removing two, and 0.502 for a different subgraph, and the cross-entropy loss increases accordingly.

Q3. We appreciate this suggestion and have added this experiment. We note that ITXGNN has the highest mean Jaccard with the ground-truth motif on all five datasets, at matched explanation size (Table R2). The margin is wide on BA-2Motif (0.994 vs at most 0.640) and Benzene (0.803 vs at most 0.545). Together with Table R1, this suggests that the method with the most stable explanations is also the most accurate with respect to the ground truth. The stability gains therefore do not come at the expense of correctness.

Faithfulness (Table R6). ITXGNN's Fid− is at most 0.003 on all five datasets: its explanation alone reproduces the prediction, as expected for a self-interpretable model. Its Fid+ is highest on BA-2Motif, SPMotif and Fluoride-C. On Benzene and Alkane-C., SubgraphX scores higher (0.428 vs 0.396; 0.389 vs 0.370) but removes between 2 and 5× more of the graph (66–68% of nodes vs 13–33%), and removing more of the graph tends to change the prediction more. Ground-truth overlap is the more direct test of relevance, and there ITXGNN is best on all five datasets (Table R2). Fidelity is reported at each method's own size, shown in Table R6(c).

AuthorsOpenReview ↗

Response to Reviewer 8qHm (2/2)

Q4. Learning qq. The assignments are free parameters for the training graphs only, qi=softmax⁡(θi)q_i = \operatorname{softmax}(\theta_i), optimized jointly with the model under the full objective. They are initialized from the training labels: one-hot plus small noise when K=CK = C, and round-robin over a class's sub-prototypes when K>CK > C, so the within-class split is learned. Validation and test graphs never enter QQ.

Prototypes are not directly optimized. At each step they are recomputed as the qq-weighted mean of the mini-batch's subgraph embeddings, μk=∑iqikzi/∑iqik\mu_k = \sum_i q_{ik} z_i / \sum_i q_{ik} (Sec. 4.2), so gradients flow to both QQ and the encoder.

For unseen graphs qq is not used. After training, prototypes are computed once over the full training set with the learned QQ. A test graph is embedded and assigned by its similarity to these prototypes (softmax for soft, argmax for hard). No test-time optimization or test information is used. In Appendix D the 10−210^{-2} "prototype" learning rate applies to the prototype layer's other parameters (the prototype-to-class matrix for K>CK > C, and the LayerNorm before cosine similarity), not to the prototypes. We will relabel it.

Q5. We thank the reviewer for the suggestion and think it is an excellent addition to support claims made in the paper. We extend the existing ablations done for BA-2Motif on MUTAG. We also add 2 new baselines; a supervised contrastive loss and a prototype regularization baseline (using a center loss) with the same backbone for both MUTAG and BA-2Motif (Tables R4 and R5).

On BA-2Motif, both baselines reach perfect accuracy, but their explanations collapse to near the whole graph (intra-class nGED 0.632). The IsoRate gap drops from 0.895 to 0.000, and the nGED ratio rises from 0.109 to 0.971. On MUTAG, ITXGNN's explanations are far more consistent in both spaces: the within-/between-class ratio is 0.412 vs 0.783–0.795 (GIN) and 0.644 vs 0.882–0.906 (ECFP). Its accuracy is equal or higher (0.937 vs 0.884–0.916). On BA-2Motif (K = C), the labels already give the right clustering, so learning it adds little and the gap to frozen clustering is small. On MUTAG (K = 4) the gap is clearer: 0.500 vs 0.412 (GIN), lower on 4/5 seeds.

Class labels alone are therefore enough for accurate prediction, but not for stable explanations. Stability comes from the clustering objective, I(Ĉ;P), on both a synthetic and a molecular dataset. Full results, with more seeds and datasets, will be included in the final version.

Reviewer_9rb3OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This work defines a new version of stability for GNN explanability and proposes a method to toward this direction. They used several standard benchmarks to evaluate their method, and experiments showed that their method is more stable than existing methods that include post-hoc methods and self-interpretable methods.

Critical strengths

Studying how to make the explanations more trustwhothy is very important for real-world deployments.

Critical weaknesses
  • Figure 1 is clearly exaggerated. I am very familiar with research in this field, the explanations generated by existing methods clearly do not match the situation presented by the authors in Figure 1. It seems that the authors deliberately constructed a scenario that never occurs in real-world applications in order to make their argument hold up.
  • Some works have been misunderstood by the authors, such as Tai et al. (2026), which study "self-consistency" rather than "consistency"; and Tai et al. (2025), which did not study the explanation of the perturbation.
  • Line 134, "a recent theoretical work" does not cite any reference, so I am not sure what the work mentioned in the text is, or how it relates to the author’s own research.
  • Paper writing should be further improved. Based on my reading, I cannot distinguish this paper from existing research on explanation stability; in my opinion, it does not go beyond existing research, and the new criteria are not necessary for GNN explanations: Eq. (2) focuses on expected value so it is very weak, and the premise that "explanations belonging to the same group should be the same" is inappropriate -- even within the same label or category, some differences in explanations should be allowed and accepted for different instances.
  • The reference cited in line 190 is incorrect.
  • Assumption 1 lacks basis and is not realistic in practical applications.
  • The authors defined several new stability metrics, and to some extent, they optimized their model in this direction, so it's not surprising that their method performs well in the final results. So I strongly suggest that the authors provide results showing the plausibility of the explanations, since all the datasets they used have human-annotated explanations.
  • The first phrase in the abstract is self-interpretable GNNs, so I assumed this paper would only focus on that. However, in the experiments, I found that the authors compared their method with several post-hoc GNN explanation methods, which is weird.
Questions for authors

Please see above.

7 comments
AuthorsOpenReview ↗

Response to Reviewer 9rb3 (1/3)

We would like to thank the reviewer for reading our paper and providing constructive feedback.

W1. In regards to Figure 1 in the paper we agree that a schematic alone does not establish the problem. The figure is a schematic illustration and we will make this more clear in the caption. However, the pattern it depicts is measured in our experiments. The intra-class IsoRate of the baselines is 0.003 (GSAT), 0.011 (ProtGNN), 0.084 (PGExplainer) and 0.097 (GNNExplainer), compared to 0.970 for ITXGNN (paper Table 1). Baseline explanations of graphs from the same class are therefore rarely isomorphic when there is a fixed planted motif. We also provide a qualitative case study in Appendix B of the paper.

W2. We thank the reviewer and will correct both descriptions in the paper and fix the citations. Tai et al. (2026) study self-consistency: whether a self-interpretable GNN produces the same explanation when its own explanation is fed back as input. They propose a fine-tuning strategy that enforces it. Tai et al. (2025) studies the consistency of explanations across models trained with different random seeds. They identify redundancy arising from weak conciseness constraints as a root cause and propose Explanation Ensemble to mitigate it. We will remove this work from the list of perturbation-based methods. In both works, the explanations compared are of the same graph. Our notion concerns agreement between explanations of different graphs (see W4).

W3. We will add the missing reference Agarwal et al. (2022). This work formalizes perturbation-based, local stability around a single instance, while we formalize stability across instances.

AuthorsOpenReview ↗

Response to Reviewer 9rb3 (2/3)

W4. We will revise the introduction and related work to make this positioning explicit. We will also revise the writing throughout, particularly the introduction, related work and definitions. The notion we study concerns a different axis of stability: whether explanations agree across different graphs that share an underlying phenomenon, which the existing notions discussed below do not measure.

Existing notions of explanation stability are per-instance: robustness to input perturbations (Agarwal et al., 2022), consistency across training runs or seeds (Zhao et al., 2023a; Tai et al., 2025), and self-consistency (Tai et al., 2026). In each, the explanations being compared are of the same graph: under a perturbation of it, from a retrained model, or from the model re-applied to its own explanation. Our notion is cross-instance: graphs governed by the same underlying phenomenon (e.g. molecules that are mutagenic because they contain a nitro group) should receive consistent explanations, while graphs governed by different phenomena should receive distinguishable ones.

Two simple examples show that these are different properties. First, take two BA-2Motif graphs of the house class. Suppose an explainer returns the house for the first graph and the house plus two attached base nodes for the second, and returns the same output under perturbation, across seeds, and when reapplied to its own explanation. It then satisfies every per-instance notion above, yet the two explanations are not isomorphic and do not identify a single structure for the class. Per-instance criteria do not register this, since they do not compare explanations of different graphs; Definition 1 does. The low intra-class IsoRate of the baselines (0.003-0.097, Table 1 in the paper) reflects this kind of cross-instance variation (Appendix B). Second, an explainer that returns the same single node for every graph also satisfies all per-instance notions, which is why per-instance criteria are paired with a separate accuracy metric. Under Definition 1, its within-class and between-class dispersions are both zero, so δ < Δ fails and the criterion itself excludes it. Beyond the definition, the paper proposes an objective, I(Ĉ;P), that targets this property, and an evaluation protocol to measure it. We have a table to highlight the distinction of the related works below.

Notion of stabilityExplanations comparedHeld fixedWhat variesHow it is obtainedRepresentative work
Perturbation robustnessSame graphModelSmall perturbations of the inputVaries by workBajaj et al. (2021); Funke et al. (2022); Agarwal et al. (2022); Fang et al. (2023); Li et al. (2024; 2025)
Cross-seed consistencySame graphInstanceModel (retrained with different seeds)Post-hoc ensemble across trained modelsTai et al. (2025)
Self-consistencySame graphInstance and modelExplainer re-applied to its own explanationAdded fine-tuning lossTai et al. (2026)
Within concept coherence (ours)Different graphsModel and concept (prototype)The instanceTraining objective, I(Ĉ;P)ITXGNN

We kindly ask if the reviewer could point us to specific works that already define or optimize cross-instance explanation stability? We would discuss them in the revision.

AuthorsOpenReview ↗

Response to Reviewer 9rb3 (3/3)

As for the point about "the premise that 'explanations belonging to the same group should be the same' is inappropriate," we agree, and this is what our formulation states: even within the same label, explanations of different instances are allowed to differ. Our wording "should be the same" was misleading, and we will replace it with "should be more similar to each other than to explanations of other groups." Definition 1 is relative. It requires only that within-group dispersion be smaller than between-group dispersion (δ < Δ), not that explanations be identical. The appropriate amount of variation depends on the data. In BA-2Motif and SPMotif, every graph of a class contains the same motif, so same-class explanations can be isomorphic. In Benzene, positive molecules share a benzene ring but differ in rings and substituents, so their explanations share a core yet vary. Definition 1 accepts both, and several explanation types within a class are supported by using more prototypes than classes (K > C, e.g., K = 4 for MUTAG).

Regarding "The new criteria are not necessary for GNN explanations: Eq. (2) focuses on expected value so it is very weak." Cross-instance stability is necessary for explanations to be trusted: if graphs that belong to a class for the same reason receive disagreeing explanations, a user cannot tell which substructure the model relies on. Both parts of Definition 1 are needed. A trivial explainer that returns the same output for every graph is perfectly stable under any per-instance notion, or any criterion that only asks same-group explanations to be close, yet it says nothing about the class. Definition 1 excludes it by also requiring separation between groups (δ < Δ fails). We agree that an expectation alone hides the spread, which is why the evaluation does not rely on averages alone. IsoRate counts individual pairs, so an intra-class IsoRate of 0.970 means 97% of same-class explanation pairs are isomorphic, and Table R1 reports the spread across seeds.

W5. We thank the reviewer. The sentence names GSAT and PGIB but cites the Gumbel-Softmax papers. We will cite Miao et al. (2022) for GSAT and Seo et al. (2024) for PGIB, and keep Jang et al. (2016) and Maddison et al. (2016) for the Gumbel-Softmax operator.

W6. Theorem 1 is intended to give insight into the mechanism: under Assumption 1, maximizing I(Ĉ;P) raises a lower bound on the separation-to-scatter ratio of explanation embeddings, which is why the objective is expected to promote cross-instance consistency. It is not a guarantee for a trained model, and our claims about stability rest on the empirical results in Section 6. We will state this scope explicitly after Theorem 1.

W7. We add Jaccard against the ground-truth motifs for all methods on BA-2Motif, SPMotif, Benzene, Alkane-Carbonyl and Fluoride-Carbonyl, at matched explanation size (Table R2). Note that MUTAG (TUDataset, 188 graphs) does not provide node-level ground-truth annotations. ITXGNN has the highest Jaccard on all five datasets (e.g., 0.994 on BA-2Motif, 0.803 on Benzene), although ground truth is never used in training. Regarding optimizing toward the metric: the training objective uses only labels and clustering, never ground truth or isomorphism. IsoRate and nGED compare explanation subgraphs structurally, and ECFP embeddings are independent of our model. In the new pairwise analysis, all methods, including ours, are grouped by predicted class.

W8. Comparing against post-hoc explainers is standard in self-interpretable GNN work (e.g., GSAT, PGIB). Since stability is a property of any explainer, including post-hoc methods shows that the issue is not specific to one family of methods.

Reviewer_5U6KOpenReview ↗
Rating
2(−1)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies explanation stability across samples in self-explainable graph neural networks. The authors argue that similar input graphs should yield consistent and stable explanations, while explanations for input graphs belonging to clearly different groups should remain distinguishable. They formalize this property through the relationship between within-group and between-group explanation distances. To promote this property, the authors propose ITXGNN, which extracts explanation subgraphs using Gumbel-Softmax node masks, encodes them with a GNN, and performs classification through soft clustering and prototype similarities. The training objective encourages both mutual information between prototype assignments and cluster assignments, and information dependence between prototype assignments and class labels. The authors provide a theoretical analysis of representation separation under an idealized two-prototype setting and evaluate explanation stability and predictive performance on six synthetic and molecular graph datasets.

Critical strengths

Explanation consistency across samples is a valuable research problem. By considering both within-group consistency and between-group discrimination, the paper helps distinguish meaningful consistency from the degenerate case of producing the same explanation for every input. This perspective complements evaluation of individual explanations and provides a clear, testable research objective.

A distinctive aspect of the method is its explicit use of clustering–prototype mutual information alongside prototype–label alignment to address explanation stability across instances.The paper provides promising empirical evidence. ITXGNN achieves strong explanation stability on the synthetic benchmarks while maintaining competitive classification accuracy and AUC. The molecular experiments also report favorable stability metrics under both ECFP and GIN embeddings.

The authors also provide component ablations, explanation visualizations, and an idealized theoretical analysis to examine the method from several perspectives. These qualifications help readers interpret the theoretical results appropriately.

Overall, the paper addresses a clearly motivated problem, presents encouraging experimental results, and has potential research value.

Critical weaknesses

The independent contributions of the two optimization objectives have not been directly verified. Table 4 primarily freezes or replaces model components, which does not sufficiently distinguish the benefits of (I(P,C^))(I(P,\hat{C})), the mutual information term between prototype assignments and cluster assignments, from those of (I(P,Y))(I(P,Y)), the information term between prototype assignments and labels. Since these two objectives are central to the method, each should be removed separately, with predictive performance and explanation stability reported under model architectures, data splits, and training budgets that are kept as consistent as possible. In particular, removing (I(P,C^))(I(P,\hat{C})) while retaining the prototype-based classification path would directly test whether the additional clustering constraint improves explanation consistency. When removing (I(P,Y))(I(P,Y)), the authors should also clarify how the prediction head and the mapping from prototypes to classes are handled so that the results can be interpreted correctly.

The within-group FID used in the molecular experiments differs from the within-group dispersion defined in Definition 1. The definition concerns the expected distance between explanations from the same group, whereas Section 6.1 randomly divides the explanations of a class into two halves, computes the FID between them, and averages over repeated splits. Even when a class contains substantially different explanation types, the two halves may still have similar means and covariances, resulting in a small FID. Thus, this metric does not directly support the conclusion that explanations within the same class are close to one another. I suggest using the existing ECFP and GIN representations to sample explanation pairs from the same and different classes, compute their average within-group and between-group distances using the same distance function, and report the resulting ratios and uncertainty estimates. The existing FID results can be retained as a supplementary distribution-level analysis.

Questions for authors

How do classification performance and explanation stability change when (I(P,C^))(I(P,\hat{C})) and (I(P,Y))(I(P,Y)) are removed separately? With the same subgraph extractor, encoder, and prototype-based classification path retained, how much additional stability does (I(P,C^))(I(P,\hat{C})) provide? In the setting without (I(P,Y))(I(P,Y)), how are class prediction and the correspondence between prototypes and classes handled?

The framework of “local evidence selection—representation encoding—prototype clustering—mutual information optimization” could, in principle, also be applied to other tasks. Could the authors further explain which specific challenges in graph explanation motivated this design? In particular, relative to existing prototype-based graph explanation methods, how does the additional clustering–prototype mutual information constraint promote structural consistency across explanation subgraphs from different samples? Clarifying this would help establish the motivation for choosing graph explanation as the setting and the method’s distinctive contribution in this setting.

5 comments
AuthorsOpenReview ↗

Response to Reviewer 5U6K

We thank the reviewer for taking the time to read our paper, their service to the research community, and their constructive comments.

We agree that removing each objective separately is the most direct test of their contributions. Evaluating the independent contributions of the two optimization objectives would provide evidence determining the individual effects of each loss component on the framework. We now have conducted additional ablations on the BA2-Motif dataset (and on MUTAG, Table R5) where we remove I(Ĉ;P) and I(P;Y) respectively. Note that our framework is end-to-end so simply removing each loss term can cause the framework to simply break we therefore conduct the ablation by removing the loss term and either replacing the components that are dependent on the missing component or simply removing it altogether when feasible. As seen in Table R3 we can see that our experiments suggest that when removing I(Ĉ;P) the model's explanation stability metrics significantly decrease (IsoRate gap 0.895 → 0.084, nGED ratio 0.109 → 0.943, accuracy unchanged at 1.000). This suggests that without the pressure from the clustering and prototypes the subgraph extractor does not feel the need to find compact and consistent explanations, at this stage the model is essentially an expressive GNN driven solely by the I(P;Y) term so prediction accuracy is paramount at the cost of explanation stability. This suggests that explanation stability is provided heavily by the MI term I(Ĉ;P).

When removing I(P;Y) the model's prediction accuracy is severely degraded (accuracy 0.544, AUC 0.599) as expected. Interestingly, the explanation stability also suffers. In this ablation we substitute labels by using k-means to generate pseudolabels. At test time, prototypes are mapped to classes by the best one-to-one assignment, and accuracy is reported as max(a, 1−a). Since labels are replaced by pseudo-labels, the framework is now driven to find a clustering solely by the prototypes and psuedolabels. This is why we see increased stability in comparison to the removal of I(Ĉ;P) but still decreased stability with respect to the full framework since the model may struggle to inherently find a stable clustering without labels.

We also agree with the reviewers concern about our prior data using FID on molecular datasets to gauge stability ratios. We take the suggestion and run an experiment using pairwise distances between explanation pairs. We report our results in Table R1 (mean ± std over 5 seeds) and note that ITXGNN still retains the lowest mean ratio on both ECFP and GIN embeddings on all molecular datasets. With a much lower intra/inter ratio on 3/4 datasets.

On graphs, an explanation is a discrete subgraph, and many different subgraphs can support the same prediction. Existing explainers often select the motif plus extra nodes that vary from graph to graph (intra-class IsoRate 0.003–0.097 on BA-2Motif). Enforcing agreement directly would require subgraph matching, which is combinatorial and not differentiable.

As for the reviewer's question on motivation; existing prototype methods (PGIB, ProtGNN) pull each explanation toward a prototype independently, so any subgraph near the prototype is accepted, whether or not it matches the subgraphs selected for other graphs. I(Ĉ;P) acts on the group: it rewards tight, well-separated clusters, so the extractor is driven to select the structure shared within each cluster. The ablation above (Table R3) and the perturbation study (Table R7) suggest that this produces structurally consistent explanations.

I thank the authors for their detailed response and additional experiments. The new ablation studies and pairwise-distance evaluations have substantially addressed most of my original concerns, and I appreciate the authors' efforts to improve the paper.

However, after reading Reviewer 9rb3's comments and further examining the related literature, I confirmed the inaccuracies in the descriptions of Tai et al. (2025, 2026). Additionally, I noticed that the paper overlooks PAGE. Although PAGE differs from the proposed method, its use of prototypes and shared explanatory patterns across instances is closely related to this work. Given that cross-instance explanation stability is one of the paper's central contributions, these issues raise further concerns about its novelty and positioning.

Therefore, I believe the authors need to provide a more systematic and rigorous discussion of related work, clearly distinguish different notions of explanation stability, and better establish the substantive differences between this work and existing studies.

Taking into account the authors' rebuttal, the other reviewers' comments, and my own further examination, although most of my original concerns have been alleviated, I remain unconvinced that the paper has sufficiently established its novelty and positioning relative to prior work. Therefore, I have decided to revise my rating from Weak Accept to Weak Reject. I encourage the authors to further strengthen the related-work discussion and more clearly articulate the distinctive contributions of this study.

AuthorsOpenReview ↗

Corrected Tai et al. descriptions, PAGE, and positioning

We thank the reviewer for their prompt response, engagement and thoroughness in their response. We agree that the submitted version described Tai et al. inaccurately and did not discuss PAGE. PAGE is a relevant work and it will be added to the paper (in addition to the other references discussed).

Corrected descriptions. Tai et al. (2025) show that self-interpretable GNNs trained with different random seeds produce inconsistent explanations, identify redundancy from weak conciseness constraints as the root cause, and mitigate it with an explanation ensemble. Tai et al. (2026a) show that faithfulness is tied to self-consistency, i.e., whether an explanation stays the same when the model is reapplied to it, and propose self-consistency fine-tuning; Tai et al. (2026b) analyze why self-inconsistency arises and propose a training-free self-denoising step. In all three, the explanations being compared are derived from the same graph.

PAGE (Shin et al., 2024). At a high level, PAGE shares several ingredients with our method. This method uses clustering of embeddings, prototypes for clusters, and the core idea of finding a common pattern amongst similar graphs is something we also leverage. However, the two works differ in an important way. Notably, this is a post-hoc method so they utilize a fixed model (fixed embedding space). This is quite different from us since we have a built in objective that guides the training of the embedding space to learn a clustering, prototypes, and a subgraph extractor all together end-to-end. Also PAGE outputs one prototype graph per cluster and does not produce or constrain the explanations of individual graphs, nor measure whether they agree across graphs (its consistency measures robustness across model hyperparameters). ITXGNN trains each graph's own explanation, the one its prediction is based on, to agree with those of graphs in the same group and differ from those of other groups.

Systematic comparison.

Line of workExplains each instanceCompares explanations of different graphsBuilt into model trainingRepresentative work
Perturbation robustnessYesNo (same graph)NoBajaj et al. (2021); Agarwal et al. (2022)
Cross-seed consistencyYesNo (same graph)NoTai et al. (2025)
Self-consistencyYesNo (same graph)Fine-tuning (2026a); post-processing (2026b)Tai et al. (2026a; 2026b)
Model-level explanationNo (class-level prototypes)No (post-hoc summaries; agreement not required)NoXGNN (Yuan et al., 2020); PAGE (Shin et al., 2024)
Multi-instance alignmentYesYes (post-hoc alignment to a reference explanation)NoGNNExplainer (Ying et al., 2019)
Prototype-based self-explainingYesNo (each against its prototype)YesProtGNN; PGIB
ITXGNN (ours)YesYes (agreement within groups, separation across)Yes

Positioning. Definition 1 compares the within-group and between-group dispersion of a model's instance-level explanations, and I(Ĉ;P) promotes this property during training by maximizing the mutual information between learned cluster assignments and prototype-similarity assignments. To our knowledge, no prior work on GNN explanation defines this property or promotes it during training; our contribution is the notion, the objective, and the metrics that measure it. The property does not follow from existing means: ITXGNN has a lower within-/between-class ratio than both prototype-based methods (ProtGNN, PGIB) in all 8 dataset representation comparisons (Table R1), and class-conditioned objectives on the same backbone are far less consistent at equal or lower accuracy (Tables R4, R5).

Additional experiments. Tai et al. (2025; 2026a) evaluate the explanation ensemble and self-consistency fine-tuning on same-graph consistency, including GSAT on Benzene (Tai et al., 2026a, Table 1). We are extending this comparison to cross-instance stability, which they were not designed or evaluated for, on MUTAG and Benzene (with ground-truth Jaccard on Benzene, as our MUTAG release has no ground-truth masks). We will post initial results in this thread by November 18 (AoE).

References.

  • Shin, Kim & Shin (2024). PAGE: Prototype-based model-level explanations for graph neural networks. IEEE TPAMI 46(10).
  • Tai, Liu, Zhong & Zhou (2026b). Why self-inconsistency arises in GNN explanations and how to exploit it. arXiv:2605.07527.
  • Yuan et al. (2020). XGNN: Towards model-level explanations of graph neural networks. KDD.

Comments on the paper

4 comments
AuthorsOpenReview ↗

General Response (1/3): Summary of New Results

We thank the reviewers for their comments. The rebuttal adds the experiments below. Tables are numbered R1–R7 and referenced in the individual responses.

  1. Pairwise explanation stability Table R1 reports the within-/between-class ratio of pairwise explanation distances in two representation spaces. ITXGNN has the lowest ratio on all four molecular datasets in both spaces (0.43–0.84). The baselines range from 0.70 to 1.05, i.e., their same-class explanations are barely closer to each other than to explanations of other classes.
  2. Correctness against ground truth In Table R2, ITXGNN has the highest mean Jaccard with the annotated motif on all five datasets with node-level annotations.
  3. Faithfulness Table R6 reports Fid+, Fid− and explanation size for all methods.
  4. Ablation of the two objectives Table R3 (BA-2Motif) and Table R5 (MUTAG) remove each objective separately.
  5. Structural perturbation of learned explanations Table R7 shows that departing from a learned explanation moves its embedding away from the class centroid and raises the training loss.
  6. Ablation on a dataset with K>CK > C and class-conditioned baselines Tables R4 (BA-2Motif) and R5 (MUTAG, K=4K = 4): at equal or higher accuracy, the full model has more stable explanations than both class-conditioned baselines.

Table R1. Explanation stability: within-/between-class ratio of mean pairwise explanation distances, L2 (↓; a ratio of 1 means same-class explanations are no closer than different-class ones). Mean ± std over 5 seeds.

MethodMUTAGBenzeneAlkane-C.Fluoride-C.
(a) ECFP (lower is better)
GNNExplainer0.801 ± 0.1180.958 ± 0.0210.950 ± 0.0580.854 ± 0.038
PGExplainer0.930 ± 0.1340.923 ± 0.0480.938 ± 0.0210.930 ± 0.072
SubgraphX0.896 ± 0.0430.923 ± 0.0220.959 ± 0.0170.903 ± 0.037
GSAT0.858 ± 0.0830.883 ± 0.0290.937 ± 0.0211.020 ± 0.012
PGIB0.883 ± 0.1580.979 ± 0.0270.954 ± 0.0310.991 ± 0.014
ProtGNN0.805 ± 0.1110.886 ± 0.0240.927 ± 0.0520.947 ± 0.042
ITXGNN (ours)0.644 ± 0.0620.635 ± 0.0400.561 ± 0.1880.839 ± 0.045
(b) GIN (lower is better)
GNNExplainer0.719 ± 0.1360.917 ± 0.0340.957 ± 0.0550.911 ± 0.088
PGExplainer0.801 ± 0.2700.831 ± 0.0880.955 ± 0.0430.982 ± 0.054
SubgraphX0.916 ± 0.0980.934 ± 0.0180.958 ± 0.0260.976 ± 0.039
GSAT0.714 ± 0.1820.790 ± 0.0350.912 ± 0.0410.979 ± 0.042
PGIB0.937 ± 0.1030.983 ± 0.0400.939 ± 0.0880.999 ± 0.015
ProtGNN0.696 ± 0.1060.882 ± 0.0220.947 ± 0.1251.052 ± 0.020
ITXGNN (ours)0.431 ± 0.0440.662 ± 0.0270.568 ± 0.1790.581 ± 0.277

Table R2. Explanation correctness: Jaccard with the ground-truth motif (↑). Mean ± std over 5 seeds.

MethodBA-2MotifSPMotif-0.5BenzeneAlkane-C.Fluoride-C.
GNNExplainer0.225 ± 0.1150.156 ± 0.0430.074 ± 0.0210.271 ± 0.1120.155 ± 0.059
PGExplainer0.167 ± 0.0000.469 ± 0.0430.507 ± 0.2390.322 ± 0.1540.428 ± 0.079
SubgraphX0.167 ± 0.0000.208 ± 0.0000.326 ± 0.0240.243 ± 0.0020.310 ± 0.014
GSAT0.640 ± 0.0290.543 ± 0.0690.545 ± 0.0180.314 ± 0.1880.376 ± 0.075
PGIB0.160 ± 0.0650.444 ± 0.0010.514 ± 0.0440.170 ± 0.0270.141 ± 0.021
ProtGNN0.286 ± 0.0020.204 ± 0.0120.491 ± 0.0160.206 ± 0.0330.301 ± 0.021
ITXGNN (ours)0.994 ± 0.0110.559 ± 0.1450.803 ± 0.0160.374 ± 0.1600.433 ± 0.021
AuthorsOpenReview ↗

General response (2/3): Tables R3–R5

Table R3. Ablation of the two objectives on BA-2Motif (K=C=2K = C = 2). 5 seeds per arm, nGED ratio = intra-/inter-class nGED (ratio of means, 95% bootstrap CI; ↓).

VariantAcc. ↑AUC ↑IsoRate gap ↑nGED ratio ↓
Full ITXGNN1.000 ± 0.0001.000 ± 0.0000.895 ± 0.1140.109 [0.010, 0.223]
I(Ĉ;P) removed1.000 ± 0.0001.000 ± 0.0000.084 ± 0.1690.943 [0.847, 0.971]
I(P;Y) removed0.544 ± 0.0280.599 ± 0.0600.176 ± 0.1420.926 [0.877, 0.962]

Table R4. Class-conditioned baselines on BA-2Motif (K = C = 2), same backbone. 5 seeds, nGED ratio = intra-/inter-class (ratio of means, 95% bootstrap CI; ↓).

VariantAccIsoRate gap ↑nGED intra ↓nGED inter ↑nGED ratio (intra/inter) ↓ [95% CI]
ITXGNN (Full)1.000 ± 0.0000.895 ± 0.1140.030 ± 0.0330.276 ± 0.1890.109 [0.010, 0.223]
Class-cond. + SupCon1.000 ± 0.0000.000 ± 0.0000.632 ± 0.0000.651 ± 0.0010.971 [0.971, 0.971]
Class-cond. + center loss1.000 ± 0.0000.000 ± 0.0000.632 ± 0.0000.651 ± 0.0020.971 [0.971, 0.971]

Mean ± std over 5 seeds; nGED ratio = mean(intra) / mean(inter) over seeds, with its 95% bootstrap CI over seeds.

AUC is 1.000 ± 0.000 for every row.

Table R5. MUTAG (K = 4, C = 2), same backbone. 5 seeds, 19 test graphs. Pairwise ratio = within-/between-class explanation distance, grouped by predicted class (↓), as in Table R1.

VariantAcc. ↑AUC ↑Ratio, ECFP ↓Ratio, GIN ↓
Full ITXGNN0.937 ± 0.0210.967 ± 0.0130.644 ± 0.0530.412 ± 0.057
Ablations
Frozen clustering0.937 ± 0.0210.967 ± 0.0170.718 ± 0.0480.500 ± 0.069
No prototype path (linear head)0.895 ± 0.0000.951 ± 0.0150.662 ± 0.0540.464 ± 0.046
I(Ĉ;P) removed0.905 ± 0.0390.956 ± 0.0280.855 ± 0.0390.759 ± 0.056
I(P;Y) removed0.842 ± 0.0000.723 ± 0.0130.613 ± 0.0730.463 ± 0.045
Class-conditioned baselines (prototypes = class means, no I(Ĉ;P))
Class-cond. + SupCon0.916 ± 0.0260.985 ± 0.0100.906 ± 0.0210.795 ± 0.030
Class-cond. + center loss0.884 ± 0.0520.951 ± 0.0170.882 ± 0.0830.783 ± 0.080
AuthorsOpenReview ↗

General response (3/3): Tables R6–R7

Table R6. Faithfulness, with each method at its own explanation size. Mean ± std over 5 seeds. Part (c) gives the explanation size behind each score.

MethodBA-2MotifSPMotif-0.5BenzeneAlkane-C.Fluoride-C.
(a) Fid+: prediction change when the explanation is removed (↑)
GNNExplainer0.219 ± 0.2190.435 ± 0.0370.265 ± 0.0500.246 ± 0.0250.252 ± 0.071
PGExplainer0.093 ± 0.0060.450 ± 0.0540.396 ± 0.1110.360 ± 0.0970.284 ± 0.032
SubgraphX0.340 ± 0.1310.539 ± 0.0000.428 ± 0.0200.389 ± 0.0190.193 ± 0.011
GSAT0.354 ± 0.0310.486 ± 0.0370.317 ± 0.0110.212 ± 0.0330.302 ± 0.081
PGIB0.143 ± 0.072-0.003 ± 0.0000.189 ± 0.0280.193 ± 0.0500.234 ± 0.121
ProtGNN0.308 ± 0.0710.433 ± 0.0120.359 ± 0.0150.272 ± 0.0460.185 ± 0.049
ITXGNN (ours)0.390 ± 0.0260.568 ± 0.1210.396 ± 0.0070.370 ± 0.1420.710 ± 0.026
(b) Fid−: prediction change when only the explanation is kept (closer to 0 is better)
GNNExplainer0.297 ± 0.1520.440 ± 0.0320.439 ± 0.0200.269 ± 0.0570.154 ± 0.007
PGExplainer0.000 ± 0.0000.272 ± 0.0740.439 ± 0.0350.189 ± 0.0530.151 ± 0.015
SubgraphX0.000 ± 0.0000.580 ± 0.0000.138 ± 0.055-0.018 ± 0.024-0.023 ± 0.010
GSAT0.519 ± 0.0270.317 ± 0.0400.343 ± 0.0070.345 ± 0.0850.309 ± 0.042
PGIB0.411 ± 0.0520.000 ± 0.0000.066 ± 0.0430.073 ± 0.1170.108 ± 0.104
ProtGNN0.069 ± 0.0020.497 ± 0.0030.000 ± 0.0090.021 ± 0.0320.030 ± 0.054
ITXGNN (ours)0.000 ± 0.0000.003 ± 0.0110.000 ± 0.0000.001 ± 0.0030.000 ± 0.000
(c) Fraction of nodes selected (size reference)
GNNExplainer0.2340.2370.2380.2310.338
PGExplainer1.0000.3220.3370.4220.337
SubgraphX1.0000.2370.6620.6750.573
GSAT0.2510.2540.2400.2350.237
PGIB0.3260.8650.4560.4740.457
ProtGNN0.4760.2370.5970.6330.610
ITXGNN (ours)0.1680.1990.3340.1340.238

Table R7. Perturbing the learned explanation of a trained, frozen model (BA-2Motif). We measure the distance of the perturbed explanation's embedding to its class centroid (normalized by prototype separation γ) and the change in cross-entropy. Mean ± std over 7 seeds.

TypeVariantJaccard to motifDist. to centroid / γΔCE
–learned explanation1.00000
Addition+1 node0.8330.001 ± 0.0010.000 ± 0.000
+2 nodes0.7140.018 ± 0.0220.033 ± 0.041
+3 nodes0.6250.031 ± 0.0150.059 ± 0.030
+5 nodes0.5000.066 ± 0.0210.122 ± 0.040
Removal−1 node0.8000.101 ± 0.0340.198 ± 0.068
−2 nodes0.6000.210 ± 0.0730.412 ± 0.143
Swapdifferent subgraph0.0000.502 ± 0.0550.976 ± 0.098
AuthorsOpenReview ↗

Summary of revisions and additional results (Table R8).

We thank all reviewers for their constructive feedback.

New experiments

  • Pairwise intra/inter explanation-distance ratios on all molecular datasets, replacing split-half FID as the primary metric (Table R1).
  • Ground-truth Jaccard on datasets with ground-truth motifs (Table R2).
  • Separate ablations of I(C^;P)I(\hat{C};P) and I(P;Y)I(P;Y), including datasets with K>CK > C.
  • Comparisons with supervised contrastive learning and an alternative prototype regularization objective in place of I(C^;P)I(\hat{C};P).
  • A structural perturbation test linking embedding geometry to explanation structure.
  • A comparison with Explanation Ensemble (Tai et al., 2025), Table R8 below.

Revisions to the manuscript (in progress)

  • A systematic comparison of stability notions in Section 2, with corrected descriptions of Tai et al. (2025, 2026) and a discussion of PAGE.
  • A clarification that Definition 1 is relative (δ<Δ\delta < \Delta) and allows multiple explanation modes per class through K>CK > C prototypes.

Key finding from Table R8. Explanation Ensemble improves same-graph agreement from 0.540 to 0.746 on MUTAG and from 0.705 to 0.856 on Benzene. Its mean cross-instance ratios also improve slightly, although ITXGNN retains lower mean ratios in all four dataset–embedding comparisons. Stronger cross-seed reproducibility can therefore coexist with weaker cross-instance agreement, which is consistent with the two being distinct, complementary properties.

Table R8. Cross-instance explanation stability of GSAT with Explanation Ensemble (EE; Tai et al., 2025; M = 10 models per ensemble) vs ITXGNN. Mean ± std over 5 seeds. Pairwise ratio = mean within-class / mean between-class explanation distance, grouped by predicted class (intra/inter, lower = more stable). Same-graph agreement = mean pairwise Jaccard of each test graph's explanation across the 5 runs (higher = more reproducible); it is reported for the GSAT variants to show the effect of EE, while ITXGNN is evaluated on cross-instance stability.

MUTAG

MethodAccAUCSame-graph agreement ↑Pairwise ratio, ECFP-L2 ↓Pairwise ratio, GIN ↓
GSAT0.800 ± 0.0510.918 ± 0.0210.5400.861 ± 0.0490.691 ± 0.047
GSAT + EE0.810 ± 0.0260.905 ± 0.0100.7460.852 ± 0.0480.632 ± 0.038
ITXGNN (ours)0.937 ± 0.0210.974 ± 0.008n/a0.644 ± 0.0620.431 ± 0.044

Benzene (1,200 test graphs)

MethodAccAUCGT Jaccard ↑Same-graph agreement ↑Pairwise ratio, ECFP-L2 ↓Pairwise ratio, GIN ↓
GSAT0.901 ± 0.0050.964 ± 0.0030.532 ± 0.0270.7050.881 ± 0.0260.797 ± 0.032
GSAT + EE0.911 ± 0.0020.973 ± 0.0010.576 ± 0.0090.8560.869 ± 0.0080.792 ± 0.005
ITXGNN (ours)0.940 ± 0.0010.973 ± 0.0030.803 ± 0.016n/a0.635 ± 0.0400.662 ± 0.027

Comparison: ITXGNN has a lower mean pairwise ratio than GSAT + EE, the stronger variant, by 0.13–0.23 on every dataset and embedding. EE reduces the mean pairwise ratios by 0.005–0.059 across the four comparisons.

Protocol: GSAT uses the Table R1 configuration, and explanations are extracted with the same top-k protocol (k = final_r = 0.5, ties included). EE averages the soft attention scores (and logits) of 10 independently trained GSAT models, as in the authors' code. Explanation sizes: MUTAG about 9–10 nodes; Benzene about 22 edges. GT Jaccard is computed over test graphs containing a benzene ring, with the same protocol as Table R2; the ITXGNN value is from Table R2. On MUTAG, two GSAT seeds fall below the pre-specified accuracy threshold (0.75); excluding them gives the same conclusion. MUTAG results are based on a small test set (19 graphs) and should be read alongside Benzene.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.