acceptodds
Under review as a conference paper at ICLR 2027

Discover, Ground, Explain: Designing Intrinsically Interpretable GNNs

Abstract

Molecular graph models can achieve high predictive performance while their attributions poorly localise the substructures underlying the target property. We introduce Graph Variational Motif Prior (GraphVMP), a variational GNN for intrinsic graph explainability that aligns a dedicated latent partition with graphlet descriptors through maximum mean discrepancy (MMD). To learn structural references without concept supervision, we design a novel unsupervised grouping model requiring only upper bounds on subgraph size and the number of distinct motif types as prior information. We also introduce MotifCausalShift, a controlled structural out-of-distribution (OOD) benchmark that varies spurious motif correlations and introduces unseen test-time distractors. On this benchmark, grouping recovers all four planted training motifs and isolates an unseen Star motif without a corresponding learned template. Without annotations specifying which motifs are causal, GraphVMP achieves 97.0 ± 1.0% accuracy and 85.0 ± 6.0% explanation ROC-AUC in the primary MotifCausalShift stress test, improving explanation ROC-AUC by 29.6 percentage points over the baseline without MMD. Following causal-motif replacement, node ROC-AUC remains 85.7 ± 13.7%, compared to 56.0 ± 19.4% without alignment. These results support structural discovery and explicit latent alignment as an approach to designing intrinsic graph explanations whose localisation remains effective under controlled structural shifts.

Primary area:
learning on graphs and other geometries & topologies
Keywords:
interpretable graph neural network

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_DjBhOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper proposes GraphVMP, an intrinsically interpretable variational GNN that aims to ground explanations in structural concepts discovered from data. It first introduces an unsupervised grouping model to learn a shared dictionary of subgraph templates without class labels or motif annotations, and uses their graphlet histograms as structural references. GraphVMP then aligns dedicated structural latent representations with these references through an MMD objective, while retaining residual latent channels for prediction. Node-level explanations are obtained by projecting structural latent activations back to the input graph.

The paper also introduces MotifCausalShift, a synthetic benchmark designed to evaluate prediction and explanation under controlled spurious motif–label correlations and structural distribution shifts. Experiments demonstrate improved explanation localization and OOD robustness compared with the model without MMD alignment, while maintaining competitive prediction accuracy on several standard graph benchmarks.

Critical strengths

S1: The paper introduces a grouping model that discovers shared subgraph templates without class labels or motif annotations, providing an interesting approach to constructing structural references for graph explanations. On MotifCausalShift, the model successfully recovers all four planted training motifs and isolates an unseen Star motif without a corresponding learned template. These results suggest that meaningful structural concepts can be discovered without explicit motif supervision.

S2: GraphVMP provides a systematic framework for aligning dedicated structural latent representations with discovered graphlet descriptors through MMD. The proposed MotifCausalShift benchmark is also useful for evaluating explanations under controlled structural shifts. In the primary stress test, GraphVMP improves node-level explanation ROC-AUC from 55.4% to 85.0% over the baseline without MMD, while achieving 97.0% classification accuracy. The improvement in localization persists after causal-motif replacement, providing encouraging evidence for the effectiveness of structural alignment.

Critical weaknesses

W1: GraphVMP aligns structural latent representations with graphlet descriptors through MMD, but this distribution-level alignment does not necessarily establish graph-specific correspondence between the latents and their underlying motifs. More importantly, the explanations are generated from the structural latents, while the classifier predicts from the full latent representation, including unconstrained residual channels. Consequently, the model could potentially achieve good structural localization while relying primarily on the residual channels for prediction. The paper does not provide sufficient evidence that the structural latents actually drive the predictive decisions, which is important for supporting its claim of intrinsic interpretability.

W2: The MMD targets are constructed from all detected templates, including spurious motifs, without specifying which structures are causally relevant. It is therefore unclear what mechanism encourages the model to prioritize causal motifs when both causal and spurious motifs are predictive during training. Furthermore, in Experiment 3, the spurious motifs are replaced at test time by an unseen Star motif, while the causal motifs remain familiar. The high explanation ROC-AUC may partly reflect the distinction between previously learned and unseen structural patterns, rather than the ability to distinguish causal from spurious motifs. Additional controlled comparisons are needed to separate these explanations and substantiate the claimed causal-subgraph localization.

Questions for authors

Q1: In Eq. (9), node-level explanations are computed from the structural latents, while classification uses the full latent representation. Could the authors provide evidence that predictions actually depend on the structural channels? For example, comparing structural-only and residual-only prediction, or intervening on the structural latent channels, would help establish whether the explanations faithfully reflect the model's predictive behavior.

Q2: In Experiment 1, causal and spurious motifs are perfectly correlated with the labels during training, while MMD alignment does not explicitly distinguish between them. Could the authors clarify which inductive biases explain GraphVMP's preference for the causal motifs? Additionally, for Experiment 3, could the authors evaluate test graphs that retain both familiar causal and familiar spurious motifs, and report their explanation scores separately? This would help determine whether the improvement comes from causal discrimination or from differences in motif familiarity.

Q3: The number of templates (M) and maximum subgraph size (k_{\max}) are important hyperparameters for the unsupervised grouping model, but their experimental values are not reported. Could the authors provide the settings for each dataset and analyze their sensitivity? In particular, since the number and size of meaningful motifs are generally unknown in real-world graphs, it would be helpful to understand how these choices affect template discovery, prediction accuracy, and explanation quality.

Q4: Section 4.2 evaluates template discovery only on the synthetic MotifCausalShift dataset. Could the authors provide qualitative or quantitative results on molecular datasets with known substructures, such as Mutagenicity, Benzene, AlkaneCarbonyl, or FluorideCarbonyl? Visualizing learned templates and comparing them with known functional groups would help demonstrate whether the proposed grouping method can discover meaningful structures in more realistic settings.

Q5: Table 2 reports single accuracy values for Experiments 1 and 2 without standard deviations or the number of runs. Given the substantial variance observed in Experiment 3, could the authors report results across multiple random seeds? Additionally, the main OOD comparison in Table 3 includes only DIR and the model without MMD. Comparisons with other representative intrinsic explanation methods, such as GSAT, CIGA, or TopInG, would help establish the relative effectiveness of GraphVMP.

Q6: A few details would benefit from clarification:

  • The metric (p_{\text{keep}}) in Table 4 is not explicitly defined.
  • The results in Table 3 and Appendix Tables 8–9 are not fully consistent, including the reported means and standard deviations for Experiment 3.
  • The compared baselines are not systematically introduced in Section 4.1, and PiGNN and KerGNN in Table 5 appear to lack corresponding references.
  • Additional details on the construction of graph-specific MMD targets, particularly how the detector handles multiple motif instances or unseen structures, would improve reproducibility.
3 comments

Thanks for invaluable review.

W1 (Grounding). (i) Tabs. 2–6’s classifier reads residuals (only 0.2% of absolute logit-margin contribution); causal channels give −7 to −8 Grid/+10 to +11 Wheel logits (Tab. 6). Revised GraphVMP bars residuals: margins exactly sum template-channel contributions plus bias (§3.2); reopening them yields seed-dependent decisions (App. E). (ii) Deleting a motif leaves ≤2.5% of its channel contribution at all sizes (Tab. 13), returning both classes to chance (P=49%/52%; Tab. 15). Each channel responds to its motif in each graph.

W2 (Causal preference/familiarity). All-motif MMD targets do not specify causality. Predictive fit alone cannot identify causal roles under perfect training co-occurrence (§3.1). Exp. 1 supplies causal-motif descriptors as a preliminary sanity check. Exp. 2/3 target both types, differing in training reliability: causal motifs always set labels; spurious ones are balanced in Exp. 2, 80% label-aligned in Exp. 3. MMD grounds structure; classifier training determines use. Tab. 4’s lower rows keep both familiar motif types. Localization still improves over no MMD (69.8–70.1% vs 56.1–56.2%). Familiarity alone cannot explain this, though gains weaken vs Star. Other interventions: spurious deletion retains 98.1% accuracy; spurious/causal swaps flip 5.3%/88.1% of predictions (Tab. 5). This supports preferential, not exclusive, causal dependence.

Q1. See W1(i) and the residual-readout ablation (App. H/Tab. 16).

Q2 (Bias/familiarity). See W2. Exp. 1’s supplied causal information differs from Exp. 2/3’s all-motif targets.

Q3. MotifCausalShift Stage 1: Kg=6K_g=6, dictionary capacity M=8M=8, k_\max=10; Stage 2 keeps four templates (Q6(d)). App. J/Tab. 17 retrains discovery: three seeds/setting. All four motifs are recovered at k_\max=10,M=6,8,12,16 and M=8,k_\max=12,15. At k_\max=8, every seed misses nine-node Grid. These test dictionary recovery. App. G tests downstream dictionary/detection errors: deleting a causal template may preserve accuracy yet raise spurious channels’ absolute share (22.7% without Grid). Thus, dictionary coverage matters for explanations even at high accuracy.

Q4. Tab. 1(b) reports held-out Benzene/Mutagenicity grouping. Recall/precision: Benzene 0.85/0.77; Mutagenicity 0.97/0.31 for NO₂/NH₂, often with the attached ring. App. I/Fig. 6 shows learned Benzene templates/held-out groups. All three seeds recover a six-membered-ring template. Discovery is topological: Stage 1 omits atom types; heterocycles also match. These dictionary results differ from Mutagenicity grouping, which does not establish isolated functional-group templates.

Q5. Tab. 2 gives eight-seed mean ± SD for Exp. 1/2. Tab. 3 adds CIGA/TopInG alongside DIR. CIGA has comparable localization (85.1±7.8% vs GraphVMP’s 85.8±6.8%) but chance accuracy (50.5±1.3%). TopInG is worse on both metrics.

Q6. (a) We rename pkeepp_{\mathrm{keep}} original-label retention: the fraction of edited graphs predicted as the original label. PP (Tabs. 13–16) instead is mean predicted original-label probability. (b) Tab. 3 uses Tabs. 11–12’s eight runs: accuracy/ROC-AUC 97.5±0.9%/85.8±6.8% with MMD, 91.0±9.9%/55.0±18.6% without. (c) §4.1 introduces baselines, cites PiGNN/KerGNN, and sources reported values. (d) App. K details detection/targets. For MotifCausalShift:

  1. We keep the four learned templates most often recovered as exact Stage 1 training groups, label-free. Binarized templates use exact induced VF2 matching, no timeout/truncation. Copies need ≤b=4b=4 external edges (generator’s largest planted boundary); all distinct copies at the minimum accepted boundary remain. All planted training copies are recovered exactly, no false positives. At b=6/8b=6/8, recall is 1.00; only House/Cycle give false positives: ≤0.8%/6.7%, respectively.

  2. Presence targets sm(G)∈{0,1}s_m(G)\in\{0,1\} ignore counts: repeats/overlaps do not increase them. Deletion training (§3.4) independently deletes each detected copy with probability 1/21/2. A copy survives only if all its nodes do. MotifCausalShift’s planted copies are disjoint: ≤one/template/graph.

  3. Unmatched structures add no template-presence entry. No detections gives zero references. Detection builds training targets/deletion copies; inference needs no detection. Unseen Star has no dictionary entry; test interventions assess its effect.

  4. H∈N6×4H\in\mathbb N^{6\times4} columns are the four templates’ 6D graphlet descriptors. Each graph has six references:

    tj(G)=(Hj1s1(G),…,Hj4s4(G))/45∈R4,j=1,…,6.t_j(G)=(H_{j1}s_1(G),\ldots,H_{j4}s_4(G))/45\in\mathbb R^4,\quad j=1,\ldots,6.

    Divisor 45, the templates’ largest descriptor entry, scales to [0,1][0,1]. These are compared to six structural posterior means μis∈R4\mu_i^s\in\mathbb R^4.

  5. Eq. (3)’s biased estimator compares six means/six references per graph, then batch-averages. The kernel is

    k(a,b)=∑α∈{0.1,1,10,50,100}exp⁡(−∥a−b∥2/(2α2)).k(a,b)=\sum_{\alpha\in\{0.1,1,10,50,100\}}\exp(-\|a-b\|^2/(2\alpha^2)).

    Matching does not pool graphs’ references.

Thank you for the detailed rebuttal and substantial revisions. I appreciate the authors' efforts to address my concerns.

The additional channel-level analyses, motif interventions, familiar-distractor evaluations, and ablation studies have significantly strengthened the empirical support for the proposed approach. Most of my original concerns have been adequately addressed. Accordingly, I have increased my Soundness score from 2 to 3.

However, I maintain my overall rating of 2 (Weak Reject), primarily because the main experimental results and the newly introduced theoretical and ablation analyses rely on different model configurations. This leaves some uncertainty about the extent to which the stronger interpretability claims are supported by the primary results. The distinction between structural grounding and causal identification also warrants careful framing.

I appreciate the improvements and do not consider additional experiments necessary at this stage. Clearer alignment between the claims and existing evidence would further strengthen the paper.

Thank you for acknowledging the improvements and increasing the soundness score.

One configuration Thank you for raising this. We reran all the experiments and now all Exp. 3 results, in the main text and the appendix come from a single configuration. We hope this resolves your concern, and we would be grateful if you would consider updating your score.

Reviewer_R1yCOpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper presents GraphVMP, a variational GNN that links structure discovery with intrinsic graph explanations. It first learns a shared subgraph dictionary without class labels or motif annotations. It then uses MMD to align structural latent components with graphlet descriptors from this dictionary. The authors evaluate prediction and explanation quality on standard graph benchmarks and MotifCausalShift, which tests controlled structural changes. The results show improved localization over the version without MMD, including after replacement of the causal motif.

Critical strengths
  1. A useful connection between structure discovery and explanation learning.

    The method learns a subgraph dictionary without class labels or motif annotations. It then uses structural descriptors from this dictionary to guide explanation learning. The recovery of planted motifs supports the discovery stage on the synthetic benchmark.

  2. Controlled tests provide evidence of improved explanation quality.

    The experiments change spurious correlations, introduce unseen distractors, and replace or delete motifs. The model with MMD improves localization over the baseline, including after replacement of the causal motif. These tests assess explanation quality beyond classification accuracy.

Critical weaknesses
  1. The key design choices need clearer justification.

    The motivation for choosing MMD and the proposed dictionary discovery method could be clearer. A more explicit discussion of their differences from related approaches would help readers understand the benefits and trade-offs of these choices.

Questions for authors
  1. Could the authors explain why MMD and the proposed dictionary discovery method are suitable for this framework?

    For example, how does the proposed dictionary discovery design differ from approaches such as MICRO-Graph, and what benefits or trade-offs does it offer?

3 comments

We thank the reviewer for recognizing the novelty of our work and for the pointer to MICRO-Graph, which we now discuss in Section 2.

Why MMD? The pooled latent tokens have no prescribed correspondence to the reference vectors. MMD compares the two sets directly: it is differentiable and invariant to token order, so it needs neither a token-to-reference assignment nor an inner matching problem, as Hungarian or optimal-transport losses would. With a characteristic kernel, it vanishes only when the token set equals the reference set; summing over tokens, each structural channel then equals its template's presence in the graph, up to a fixed scale. We apply it separately to each graph, between the structural posterior means and that graph's references, built from normalized graphlet descriptors, with a multi-bandwidth Gaussian kernel (Section 3.2; Appendix K). Removing it lowers explanation ROC-AUC from 85.8 ± 6.8% to 55.0 ± 18.6% (Table 3); we have not compared it with other set losses.

Why this discovery design, and how does it differ from MICRO-Graph? Both learn shared motifs without motif annotations by grouping nodes, but they differ in representation and purpose:

MICRO-GraphGraphVMP
Motif representationPrototype embeddings of subgraphsTemplate adjacency matrices
Discovery objectiveEM clustering of subgraph embeddings, with contrastive learningGW matching of node groups to templates, with adjacency reconstruction and group refinement
Node featuresUsedNot used (structure only)
Use of motifsGuide subgraph sampling for contrastive pre-trainingGraphlet references for grounding explanatory latent components

MICRO-Graph uses node features and embeddings to learn motifs on large molecular datasets. Its prototypes are visualised through nearest subgraphs; the embeddings themselves do not directly define discrete templates for exact detection or structural edits.

Adjacency-valued templates suit our objective: they can be inspected directly, compared with node groups by GW without a node correspondence, and converted into permutation-invariant graphlet descriptors. Reconstruction ties groups and templates to the observed edges rather than to embedding similarity alone (Section 3.3). All four planted templates are recovered exactly in every seed (Table 1), and for every tested dictionary size and every size bound that admits the nine-node Grid (Appendix J). On Benzene, the six-membered ring is a template in all three seeds (Appendix I).

Thank you for the detailed responses and additional experiments. The clarifications and visualizations make the method easier to understand. I also appreciate the authors’ efforts to address the other reviewers’ concerns, particularly the relationship between the structural explanations and the model’s predictions.

However, the motivation behind some design choices, particularly the use of MMD and the dictionary discovery approach, is still not entirely clear to me.

Thank you for your time, effort, and your valuable feedback.

Why MMD?

  • We want explanatory channels with a structural meaning that can be checked independently of the classifier.
  • This motivates explicit adjacency templates: we can inspect them, detect their occurrences and use them for interventions. GW handles unknown node correspondence; reconstruction ties discovery to observed edges.
  • MMD then aligns unordered latent tokens with each graph’s structural references without requiring a token-to-reference assignment. With our characteristic kernel, zero MMD implies matching multisets, so summing the tokens gives scaled template presence. That is the connection to interpretability.

We should also acknowledge that MMD is a suitable choice. It’s not uniquely necessary. We haven’t compared it against other set losses, so we cannot say it’s superior.

Why this discovery design, and how does it differ from MICRO-Graph? Both approaches learn shared motifs without motif annotations but differ in representation and purpose. Our explicit adjacency templates can be inspected and checked by isomorphism: all four planted motifs are recovered in every seed across the tested dictionary sizes and size bounds that accommodate the nine-node Grid (Table 1; Appendix J). Exact matching in new graphs supplies per-graph structural targets and identifies instances for controlled motif interventions. MICRO-Graph’s prototype embeddings do not themselves define discrete adjacency templates.

Reviewer_9yaPOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper introduces GraphVMP, a graph neural network designed to improve interpretability. Its main contribution is a variational GNN whose explanatory latent components are structurally grounded. It explains its own predictions by pointing to the subgraphs/motifs that actually caused its decision, rather than relying on post-hoc attribution. GraphVMP is evaluated on multiple benchmarks.

Critical strengths
  1. The proposed method directly engages with the non-identifiability problem that undermines most self-explaining GNNs, which is valuable. Also, the idea of unsupervised subgraph discovery seems interesting.

  2. The authors have provides experiments in different aspects, which is nice. The controlled experiments are useful. The reported improvements from the method seem substantial.

Critical weaknesses
  1. Every benchmark in the evaluation protocol uses small graphs. The scalability and the runtime aspect of this method are untested.

  2. There are no clear ablations provided to separate the contributions of the different components. There are many components, such as the unsupervised discovery stage, the MMD grounding, and the transformer-based pooling. They are all tested together, so it is difficult to tell which one drives the reported gains.

  3. The proposed architecture and experiments do not show that the grounded latent components actually drive the classifier's predictions. A theoretical link between structural alignment and prediction faithfulness would help clarify when the highlighted motifs really explain the model's decisions. Overall, all claims of this paper are empirical. The paper would be stronger if the authors added theoretical grounding for their central claims.

  4. The presentation in the paper requires improvement. Appendix organization is incomplete. For example, Appendix C refers to tables “below” that appear earlier, while Appendix D contains only a heading. This makes the supporting results difficult to find. Also, several acronyms and terms are never defined. One example is WL (Weisfeiler–Lehman), which appears in Section 3.3 and in Algorithm 1. Readers familiar with GNN expressiveness literature will recognize WL, but the paper should not assume this.

  5. GraphVMP method has two stages, and the second stage seems cannot correct errors from the first. The paper does not test how the method behaves when the learned motif dictionary is incomplete or corrupted. As a result, the reported gains may depend on near-perfect motif discovery, which is a condition that may not hold on real, noisy graphs.

Questions for authors

The paper has a meaningful idea, but there are some substantial issues as pointed out in the weaknesses section. I would like to request that the authors refer to that section.

3 comments

We thank the reviewer for approving our novelty, and the detailed assessment. We address all five concerns below and respectfully ask them to reconsider their assessment and score in light of the new evidence.

W1 — Scalability and runtime. Appendix E tests both stages on larger graphs. Stage 2 achieves 96.9–99.5% Star-test accuracy across 40–250 nodes with the 40-node dictionary (Table 13). Retraining Stage 1 at 250 nodes recovers all four motifs exactly in all three seeds in ~7 minutes per seed on a laptop CPU (Table 14), using one unmatched background group. Appendix F derives costs, including quadratic graph-size terms and one-off cubic Laplacian preprocessing. Detection constructs training targets and is not required for predictor inference.

W2 — Component ablations. Removing MMD lowers explanation ROC-AUC from 85.8 ± 6.8% to 55.0 ± 18.6% and accuracy from 97.5 ± 0.9% to 91.0 ± 9.9% (Table 3); Tables 11–12 report the weight sweep.

Appendix H, Table 16 changes one component at a time over three seeds. Replacing learned templates and detections with planted motifs yields identical Stage 2 weights in these runs. Removing deletion training lowers Star accuracy from 96.9% to 86.3%. Mean pooling lowers Star accuracy to 88.3% and reversed-spurious accuracy from 99.3% to 82.6%, while increasing MMD loss from 0.04–0.06 to 0.58–0.61. The residual-readout ablation is also reported. Appendix H distinguishes this revised configuration from the original main-text MMD comparisons.

W3 — Theoretical link between alignment and decisions. Appendix L now provides propositions, proofs and an empirical check. For the revised structural-only affine head, the margin decomposes exactly as Δ(G)=∑mwmcm(G)+b\Delta(G)=\sum_m w_m c_m(G)+b, where cmc_m sums the structural posterior means. Each wmcm(G)w_m c_m(G) is the channel’s Shapley value under zero channel masking (Proposition 1).

Our characteristic Gaussian MMD vanishes only when the six token means and six references coincide as multisets. Hence cm(G)=amsm(G)c_m(G)=a_m s_m(G), where sms_m denotes detected presence and am=∥h(Bm)∥1/45a_m=\|h(B_m)\|_1/45 is fixed by the template’s graphlet descriptor (Proposition 2). Let G′G' delete every instance of motif mm while preserving other detected presences. With r=c−a⊙sr=c-a\odot s, Corollary 1 gives:

\leq \|w\|_2\bigl(\|r(G)\|_2+\|r(G')\|_2\bigr).$$ A small residual bound guarantees closeness. Proposition 3 excludes the specified absence shortcut under exact alignment on originals and deletion copies. Proposition 4 characterises channel selection for fixed, exactly aligned representations. We test this link on 500 Exp. 3 graphs across three revised-model seeds. The per-graph contribution $w_m c_m(G)$ matches the causal-motif deletion effect with normalized mean absolute errors of 4% for Grid and 1% for Wheel (mean absolute error divided by mean absolute effect; Fig. 7). The ideal constant $w_m a_m$ misses the deletion effect by 25% on average under the unseen Star; exact alignment is not assumed empirically. The main-text interventions separately show 88.1% versus 5.3% prediction-flip rates after causal versus spurious swaps, and 53.4% versus 98.1% original-label retention after their deletion (Table 5). After causal swaps, node ROC-AUC remains 85.7 ± 13.7% with MMD versus 56.0 ± 19.4% without (Table 4). On Benzene, ring deletion lowers mean predicted benzene probability from 99.4% to 0.0%, versus 74.2% after atom-count-matched non-ring deletion (single run; Table 7). **W4 — Presentation.** We reorganized the appendices and added discovery examples, scaling, dictionary-error studies and ablations. WL (Weisfeiler–Lehman) and its hashing procedure are explained at first use in Section 3.3. GCN, GIN and BCE are expanded, and the broken section reference is removed. **W5 — Errors from Stage 1.** Stage 2 cannot repair its fixed dictionary. Appendix G, Table 15 tests incomplete dictionaries and noisy detections over three seeds. We remove individual causal/non-causal templates, inject false detections with probability 0.25 per absent template and graph, and independently omit occurrences with probabilities 25%, 50% or 75%, fixed before training. Perturbations precede construction of both MMD targets and deletion copies. A missed occurrence remains in the graph but is excluded from its MMD target and is never selected for deletion. False detections yield 96.3% accuracy versus 96.9% with the full dictionary; removing House/Cycle yields 94.9%/92.8%. Missing Grid/Wheel raises the spurious contribution share to 22.7%/11.7%. Missed-detection rates of 25%/50%/75% reduce accuracy to 81.5%/73.3%/61.9%: undetected motifs can remain in deletion copies assigned uniform targets, creating conflicting supervision. The full label-free pipeline achieves 95.0% accuracy over eight seeds. These tests quantify both tolerance to specific errors and dependence on causal-template coverage and detection recall.

Thank you for the detailed response and the additional experiments. They address several of my concerns, and I have raised my soundness score. However, I am keeping my overall rating of 2.

The main results and the new analyses appear to rely on different model configurations, so the headline claims are not fully supported by the revised method. Part of the reported localisation gain also seems to depend on the distractor being unseen, and some reported results, such as the Mutagenicity explanation score, need clarification. Finally, scalability is still demonstrated only on small synthetic graphs.

Thank you for your valuable feedback.

One configuration We reran all the experiments and now all Exp. 3 results, in the main text and the appendix come from a single configuration. We hope this resolves your concern, and we would be grateful if you would consider updating your score. Regarding the rest of the comments.

1. Configurations. The abstract and Tables 3–6 now report the same revised Exp. 3 configuration: structural-only readout, motif-deletion training, and a label-free dictionary. Models with and without MMD are evaluated over the same eight seeds using identical test graphs and evaluation scripts. MMD achieves 95.0 ± 3.6% accuracy and 87.7 ± 9.4% node ROC-AUC, versus 82.1 ± 17.5% and 59.2 ± 14.3% without MMD.

2. Unseen distractor. The localization advantage persists when Star is replaced by familiar spurious motifs, both with the training association preserved (75.8 ± 9.5% vs. 53.6 ± 9.4%) and reversed (75.6 ± 9.5% vs. 54.0 ± 9.3%; Table 4). Eq. 9 projects structural activations without classifier weights, so a highlighted motif need not contribute strongly to the prediction. Table 6 shows that the causal channel has the mean margin contribution of greatest magnitude in every listed configuration. With both familiar spurious and Star distractors present, causal swaps change 91.1% of predictions, versus 5.2% for spurious swaps (Table 5), supporting preferential causal dependence.

3. Mutagenicity. The previous 100.0±0.0%100.0 ± 0.0\% explanation ROC-AUC used NO₂/NH₂ annotations to construct MMD training targets and was therefore annotation-supervised. We have replaced it in the main comparison with the label-free result: 29.6±24.4%29.6 ± 24.4\% node ROC-AUC (Table 9) and 64.1±6.5%64.1 ± 6.5\% classification accuracy (Table 8). With kmax⁡=20k_{\max}=20, Stage 1 achieves 0.98±0.000.98 ± 0.00 recall and 0.62±0.030.62 ± 0.03 precision (Table 1), but this grouping quality does not ensure useful Stage 2 targets. The selected pattern is a frequent ring fragment rather than the annotated NO₂/NH₂ units. We explicitly report this limitation of the current dictionary-selection procedure.

Scalability. We now report CPU costs for both stages on real molecular datasets (Appendix F). On Mutagenicity (4,337 graphs, up to 417 nodes), Stage 1 on 3,470 training graphs takes 69.79 minutes with 4.27 GiB peak resident memory. Stage 2 takes 4.70 ± 0.06 minutes, with a maximum peak memory of 0.648 GiB across three seeds. On Benzene (12,000 graphs), Stage 1 on 800 training graphs reaches its final training epoch in 13.62 ± 4.10 minutes; this measures elapsed time through that epoch, not the complete process. The Stage 2 process takes 8.88 minutes with 0.71 GiB peak memory. Both Stage 2 measurements include target construction and evaluation. Inference requires only a Stage 2 forward pass, without repeating discovery, template matching, target construction or reconstruction.

Reviewer_BErTOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The authors propose GraphVMP to find more accurate explanations when causal and spurious motifs occur. The method discovers structural motifs through unsupervised node grouping and learns shared motif templates. Based on the found patterns, GraphVMP uses the learned representations for graph classification while providing motif-based explanations.

Critical strengths
  1. The addressing assumption is interesting, as the accuracy of explanations is not the only metric for evaluating their quality, and when causal and spurious motifs occur, the quality of explanations can be affected.
  2. Discovering important structural patterns has been an interesting and important research topic.
Critical weaknesses
  1. The overall representation in the method is pretty poor, with a lack of justification and motivation. The (only) technical description of the proposed models is not convincing enough to support the authors' addressing of the research questions. Some notations omit definitions, which decreases the readability.
  2. The general approach to the experimental settings was quite reasonable; however, the setting is not consistent regarding the baselines to be compared. In Tables 2 and 3, the authors did not describe what the baseline is, and Tables 5 and 6 have different methods of comparison.
  3. The visualization and qualitative analysis of the found structural motif are not attached, even though it is part of the core proposed method.
  4. As the authors said, the found structural motifs can be disconnected. It is not convincing that their method and approach can thoroughly identify causal motifs, since, at least in the proposed experimental setting, causal graphs are connected.
Questions for authors
  1. How does the reconstruction loss contribute to discovering the structural patterns? Furthermore, could you justify your loss functions that you did not describe or explain in your paper?
  2. Could you add the baselines including post-hoc explanations methods at least such as GNNExplainer and PGExplainer? + Hopefully, compare the consistent baselines for experiments.
  3. If you add subgraph-level accuracy to evaluate how your explanations and ground truths are matched in the real-world dataset, it might be more helpful.
1 comment

We thank the reviewer for highlighting the importance of structural discovery and for their constructive feedback. We have addressed each concern below with clarifications, additional results, and revisions. Based on this evidence, we respectfully ask the reviewer to reconsider their assessment and score.

W1, Q1 — Motivation & losses. Predictive accuracy alone cannot identify which structure an explanation should point to (Sec. 3.1). Stage 1 thus learns label-free structural references; Stage 2 grounds its explanatory channels in them. Sec. 3 motivates each choice: GW needs no group–template node correspondence, graphlet descriptors are permutation-invariant, and MMD needs no token-to-instance assignment.

Reconstruction ties discovery to the graph: the decoder places matched templates and adds a background model (Eq. 5). BCE penalizes missing and false edges; node-wise placement errors refine groups (Eq. 6; Fig. 2). Sec. 3.3 now defines LgroupL_{\mathrm{group}}. LfitL_{\mathrm{fit}} pulls eligible groups toward templates in GW distance; normalized cut favors densely connected groups; orthogonality discourages all nodes from joining one group; and the size penalty enforces kmax⁡k_{\max}. In Stage 2, LclsL_{\mathrm{cls}} trains prediction and MMD aligns structural channels with references. KL regularizes residual posteriors toward a learned Gaussian mixture and structural posteriors toward N(0,0.12)\mathcal{N}(0,0.1^2). Deletion training encourages each channel to encode its own motif’s presence rather than another’s absence (Sec. 3.4). Notation clashes resolved; symbols defined at first use.

W2, Q2 — Baselines. “Baseline” in the submission’s Tables 2–3 was GraphVMP trained identically with γ=0\gamma=0. Sec. 4.1 now defines it, and the tables label it GraphVMP w/o MMD, isolating the alignment term. Experiment 3 also compares DIR, CIGA and TopInG, trained on MotifCausalShift (Table 3).

The standard-benchmark tables (formerly Tables 5–6, now 8–9) report published accuracies from Ragno et al. (2026) and Xin et al. (2025), and explanation ROC-AUC from Xin et al. (2025) (Sec. 4.1). We also added GNNExplainer and PGExplainer from Xin et al. to Table 9.

W3 — Visualization of discovered motifs. The revision includes four views. Fig. 2 (Sec. 4.2) follows grouping, matching, placement, reconstruction and refinement on a held-out graph. Fig. 4 (Appendix D) shows each seed’s dictionary: all four planted training motifs are exact templates in all three seeds. Fig. 5 (Appendix D) shows groups on fixed held-out graphs, including missed and extra motif nodes. Fig. 6 (Appendix I) shows the Benzene dictionary and groups on the first four test molecules per class. These show exact recoveries and imperfect boundaries. Sec. 4.2 now cites Figs. 4–6.

W4 — Connectivity and causal motifs. Appendix A’s disconnected structures refer to separate benzene rings in one molecule, treated as separate explanations, rather than a disconnected discovered motif. Normalized cut encourages but does not enforce group connectivity. We therefore measure recovery: all four connected planted motifs are exact templates in every seed (Table 1); instance-level recall and precision expose incomplete or oversized groups. Every template in Figs. 4 and 6 is connected.

Stage 1 does not identify causal motifs; it discovers recurring structures without labels. Stage 2 learns which support prediction, tested through interventions. In Experiment 3, targets cover all four motifs without marking causal ones. Deleting the causal motif lowers original-label retention to 53.4%; deleting the spurious motif leaves accuracy at 98.1% (Table 5). After a causal swap, node ROC-AUC remains 85.7% with MMD versus 56.0% without (Table 4), localizing the causal motif against background and unseen distractors.

Q3 — Subgraph-level accuracy on real data. We thank the reviewer. Table 1(b) evaluates discovered groups; we now also evaluate explanations on all 1,200 positive held-out Benzene molecules across three seeds. Following Agarwal et al. (2023), we measure node-set IoU against the best-matching provided ground-truth mask, including multi-ring masks.

The top-template explanation returns the union of instances of the detected template contributing most positively to the predicted class, without a node-score threshold. It achieves 0.883 ± 0.005 mean IoU, 99.8 ± 0.0% with IoU ≥ 0.5, and 29.0 ± 3.5% exact matches.

For the Eq. 9 node map, we select the top-∣T∣|T| nodes for each reference mask TT and take the best IoU. This yields 0.839 ± 0.066 mean IoU, 98.9 ± 0.5% with IoU ≥ 0.5, and 47.0 ± 17.8% exact matches. Random scores under the same protocol yield 0.236, 8.3%, and 0.0%, respectively, using one fixed realization across checkpoints. This protocol uses ground-truth mask sizes. Appendix M (Table 18) provides the protocol, results and visualizations.

Comments on the paper

1 comment

We thank all reviewers for their constructive feedback and for confirming the novelty of our contribution. Our revised manuscript and individual responses address the concerns raised through additional experiments, theoretical analysis, and clearer presentation. We respectfully ask the reviewers to reconsider their assessments in light of these responses and revisions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.