acceptodds
Under review as a conference paper at ICLR 2027

CIExplainer++: Causal and Interpretable Explanations for Graph Neural Networks

Abstract

Explainable Artificial Intelligence aims to make black-box models more trustworthy by presenting, in a human-understandable manner, the elements that lead to the model's output. This involves both (i) identifying components and connections with genuine causal effect on outputs and (ii) translating such structures into an interpretable representation. For the former, we introduce CIExplainer, a novel perturbation-based method grounded in causal inference for explaining Graph Neural Networks (GNNs). CIExplainer identifies the subgraph with the highest causal effects on GNN predictions using the Potential Outcome Framework. We evaluate and compare CIExplainer on various GNN architectures (GCN, GraphSAGE, GAT, GIN) and datasets. To bridge subgraph explanations with human interpretability, we further propose G2TeXplainer, a method that transforms causal subgraphs into natural language explanations that capture both feature-level and relational information.

Primary area:
interpretability and explainable AI
Keywords:
Graph Neural Networks, Explanation Method, Potential Outcome Framework, GNN Explanation, Textual Explanation

Reviews

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

This paper proposes CIExplainer++, a causal explanation framework for GNNs. It targets three problems in existing methods: reliance on data correlations, limited insight into underlying causal mechanisms, and subgraphs that are hard for non-experts to interpret. The framework include 3 modules. CIExplainer uses a potential outcome model to intervene on node features, estimate the model-level causal effect, and select the nodes with the strongest effects to form a causal explanation subgraph. It supports node classification, graph classification, and link prediction. G2TeXplainer converts the structured subgraph into natural language with 4 progressive prompting strategies; The third module combines LLM-as-judge with human evaluation to measure explanation quality across multiple dimensions. Experiments use 4 standard datasets, 4 GNNs, and 7 competing methods. CIExplainer leads on node classification accuracy and inference speed and remains competitive on graph classification. G2TeXplainer with P4 performs best in both automatic and human evaluation. Ablations support the design choices behind max aggregation and the Gaussian mixture counterfactual distribution. The authors also note limitations related to sampling range, distribution assumptions, and evaluation bias.

Critical strengths

The paper brings the potential outcome framework into GNN interpretation and replaces correlation-based attribution with causal intervention. CIExplainer does not simply identify nodes or edges associated with a prediction,it intervenes on node features, estimates the model-level effect on the output, and uses that effect to select an explanation subgraph.The approach is more principled than traditional methods and can be extended to node classification, graph classification, and link prediction.

Critical weaknesses

The paper’s core component CIExplainer defines the node feature intervention effect as a causal effect within the potential outcome framework. This framing runs into a problem with GNN message passing: changing one node’s features changes its representation, which then propagates layer by layer and can affect neighboring nodes and the target node. The target outcome therefore depends on both its own treatment and the treatment of other nodes. The paper tests temporal stability and causal transience, but it still treats the target node’s K-hop computation graph as a closed unit, assuming that outside nodes remain unaffected and that units do not interfere. That assumption depends on the neighborhood sampling range and global graph connectivity, yet the paper provides no rigorous proof or multi-scenario empirical validation.

Questions for authors

1.Given that message passing can propagate an intervention beyond the target node’s K-hop computation graph(As mentioned in weakness), how do the authors justify treating that K-hop graph as a closed unit satisfying SUTVA/no-interferences?

2.Given that G2TeXplainer and LLM-as-judge build on existing work, and the authors also acknowledge limitations of LLMs, should these two modules be repositioned as auxiliary components rather than co-equal contributions alongside CIExplainer? The current version presents them as equally important, which may obscure the paper’s core innovation.

2 comments

We thank the reviewer for identifying these points. We have revised the corresponding explanations and expect to upload a consolidated manuscript revision on Thursday, 12 November, incorporating the additional experimental results and marking changes in blue.

  1. Computation graph and SUTVA

Thank you for raising this point. We agree that the conditions supporting the computation-graph restriction should be explicit. We have revised the text to distinguish sampling feature values from neighborhood sampling and to clarify that the neighborhood remains fixed throughout the counterfactual evaluations.

Our unit is the explanation instance (G,v,w,i)(G,v,w,i) defined in Section 3.3. The treatment replaces feature ii of node ww, and the outcome is the target-class score at node vv. The original and intervened outcomes are evaluated with the same trained GNN and graph structure, with all remaining input features held fixed. Message passing between nodes within this instance is part of the mechanism through which the intervention affects the target outcome.

For the node-classification GNNs considered, the KK-hop neighborhood is determined by the number of local message-passing layers. The computation graph includes the feature dependencies within this receptive field. The neighborhood is selected once and held fixed throughout the counterfactual evaluations. Feature interventions change the messages along existing paths without adding edges or message-passing layers. Although an intervention can affect other node predictions, this does not introduce additional feature dependencies into the target’s KK layer computation.

We have revised the text to clarify these conditions. In particular, the sampling limitation discussed in the paper concerns sampling replacement feature values. The neighborhood is not resampled between counterfactual evaluations. We have also stated that restricting the computation to a neighborhood requires knowledge of the GNN’s receptive field.

  1. Position of G2TeXplainer and LLM-as-judge

We agree that CIExplainer is the main methodological contribution. We have revised the introduction to present G2TeXplainer as a secondary component that produces natural language descriptions of the explanation subgraphs, with the aim of making the selected nodes and their relationships easier to interpret.

LLM-as-judge is presented as an evaluation procedure, complemented by human evaluation. Its scores assess the generated descriptions under the stated criteria. We have revised the presentation to distinguish this evaluation procedure from the methodological contribution of CIExplainer.


We would be happy to address any further questions.

AuthorsOpenReview ↗

Thank you for your Review

Thank you again for your review. Following our reply of 10 November, we have uploaded the revised manuscript, with changes marked in blue.

The method section now explains what is held fixed during feature interventions and the conditions under which the K-hop computation graph contains the dependencies of the target prediction. The introduction also presents CIExplainer as the main method, G2TeXplainer as a secondary component, and LLM-as-judge as part of the evaluation.

Please let us know if either point still needs clarification.

Reviewer_KUM7OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

CIExplainer is a perturbation-based, instance-level GNN explainer. For a target prediction, it enumerates every node ww in the KK-hop computation graph and every feature ii, replaces xw,ix_{w,i} with a value drawn from the (marginal, training-set) feature distribution, and records the change in the predicted target-class probability, τw,i(a)=Yt(a)−Yc\tau_{w,i}(a) = Y_{t(a)} - Y_c. Scores are aggregated by a max over B=10B{=}10 draws and then a max over features; the top-ll nodes form the explanation subgraph. The authors frame this difference as a unit-level causal effect under the Neyman–Rubin Potential Outcome Framework. G2TeXplainer verbalizes the explanation subgraph with LLaMA-3-8B under four prompt variants (P1–P4).

Critical strengths
  1. The method needs only forward passes (∣VNe∣⋅d⋅B|V_{Ne}| \cdot d \cdot B of them). On the node-classification benchmarks, it is the most stable method across all four backbones (Table 2), whereas PGExplainer and GNNExplainer collapse for at least one (dataset, backbone) pair. Stability across backbones is a property practitioners care about, and it is consequential to my assessment.

  2. Aggregation (max/mean/median), sampling budget (1–100), intervention distribution (GMM/joint-GMM/Gaussian/t/uniform), and marginal-vs-joint sampling are each isolated with repeated runs. It lets a reader see exactly which design choices carry the result.

Critical weaknesses
  1. The injected node features encode the ground-truth motif, so the node-classification results measure label leakage. BA-Shapes and Tree-Grid are featureless by design; the paper adds x∼N(μc,2)x \sim \mathcal{N}(\mu_c, 2) per node class (Sec. 4), but node class there is motif membership, so each motif node carries its own answer in its features. CIExplainer perturbs features of one node at a time and does not touch edges, so it cannot detect structure. The default intervention (Alg. 3) drops the GMM component of the observed value, which under class-conditional Gaussians is a class flip. The results confirm it: on Tree-Grid/GIN, the model is at 0.59 accuracy (Table 7) yet CIExplainer scores 0.84 IoU (Table 2), contradicting the paper's own claim that explanation quality is bounded by model quality. On MUTAG, the only dataset with real features, CIExplainer matches the random explainer on three of four backbones (Table 4).

  2. The Potential Outcome Framework exists because one potential outcome per unit is unobservable. Sections 3.2–3.3 (temporal stability, causal transience) restate that Φ\Phi is deterministic. What remains is a one-at-a-time finite-difference sensitivity, which is the same quantity occlusion/feature-ablation explainers compute. So correlation vs. causation (Sec. 1) is not a distinction between this method and the perturbation baselines. The finite-budget estimand (Eq. 3) is the expectation of an extreme-value statistic and, more importantly, the paper gives two different definitions: Eq. 3 and the empirical score take a signed max over draws, while Alg. 1 takes max⁡i∣CEi∣\max_i |CE_i|. If the signed max is implemented, a node whose every perturbation lowers the target probability receives a score near zero after the absolute value.

  3. P4 feeds the LLM the predicted label, the motif definitions, and pre-computed structure detections (STRUCTURES DETECTED IN SUBGRAPH, A.9). The judge's structure criterion then checks whether the text names those same structures, and node fidelity is scored against the ground-truth subgraph (A.7), not the GNN's behavior. The 89% pairwise win-rate measures prompt-constraint compliance. Human factual consistency is non-monotone (P2 = P4 = 4.333, P3 = 3.095). A.8 says "GraphXAIN achieves slightly higher scores on the full test set" and attributes this to length bias, but Table 9 shows P4 higher on fidelity and structure and lower only on clarity, while the main-text Table 3 reports the opposite ordering on a 20-example subset.

Questions for authors
  1. Which of the signed max⁡bτ\max_b \tau of Eq. 3 or max⁡i∣CEi∣\max_i |CE_i| of Alg. 1? Please report Table 2 under max⁡b∣τ∣\max_b |\tau| as well.

  2. Please rerun BA-Shapes/Tree-Grid with class-independent features (e.g., x∼N(0,2)x \sim \mathcal{N}(0,2) for all nodes) and with the original constant features.

  3. What are the μc\mu_c? For BA-2motifs, what distribution does each node receive?

  4. What is the explanation size ll, and how were edge-mask baselines (PGExplainer, GNNExplainer) converted to node sets?

  5. InputXGradient and GNNExplainer on Tree-Grid/GIN are identical, and PGExplainer on Tree-Grid/GraphSAGE has IoU = Pr, which requires zero false negatives. Are these correct?

3 comments
AuthorsOpenReview ↗

Thank you for your Review

We thank the reviewer for the detailed assessment and for recognizing the stability of the node-classification results and the value of the ablations. Your comments helped us identify a BA-2Motifs feature-generation bug and GMM convergence failures, and correct inconsistencies in the evaluation. Thank you! Below we address the feature construction, intervention-effect formulation, evaluation protocol, and textual descriptions. CIExplainer is the main contribution; G2TeXplainer is a secondary component. Revisions are marked in blue in the paper.

AuthorsOpenReview ↗

1. Feature construction, intervention effects, and evaluation protocol (Q1–Q5)

Q1. Absolute-effect aggregation. We have corrected the text to clarify that the aggregation functions include an absolute value operator. We use absolute value because interventions that consistently lower the target-class output can have large effects. . The reported results therefore already use the requested maximum-absolute-effect score. We corrected the accompanying definitions, and Alg. 1.

We use potential outcomes to define these changes in the trained GNN. Both outputs are obtained by running the same model. Perturbation baselines can also measure effects on the model’s output. Our feature interventions leave the edges fixed, so they do not measure the effects of deleting or rewiring edges.

Q2. Non-informative features. We added an experiment in which the two informative features are retained and two non-informative (NI) features are sampled from N(0,3)N(0,3), using the same distribution for every class. We retrained the backbones for this setting. The updated BA-Shapes IoU results are:

MethodBA-shapes/GCNBA-shapes+NI/GCNBA-shapes/GATBA-shapes+NI/GATBA-shapes/GINBA-shapes+NI/GINBA-shapes/GraphSAGEBA-shapes+NI/GraphSAGE
IntegratedGradients0.6590±.000.7537±.000.7902±.000.8685±.000.5961±.000.6213±.000.8424±.000.8912±.00
InputXGradient0.7021±.000.8322±.000.8299±.000.8490±.000.5848±.000.6908±.000.8424±.000.9070±.00
Shapley Values0.6107±.030.7769±.030.7249±.030.7831±.020.5228±.030.6024±.020.7370±.020.7978±.02
PGExplainer0.8163±.050.6391±.030.5731±.050.6008±.030.8566±.030.8180±.040.6378±.140.6008±.01
GNNExplainer0.5683±.020.5864±.020.2922±.010.5119±.010.5372±.030.4519±.020.4487±.030.4222±.02
SubgraphX0.5363±.010.4395±.020.5935±.030.5167±.020.7369±.020.6569±.020.6036±.020.4705±.01
CIExplainer (ours)0.9921±.020.7395±.020.9794±.010.8746±.020.8747±.020.7001±.030.9643±.010.9127±.01

CIExplainer’s IoU decreases with the added features for all four backbones. It still has the highest mean IoU for GAT and GraphSAGE. Because the backbones were retrained, the comparison involves different fitted models.

Q3. Feature distributions. The informative feature vectors are sampled from N(μc,2I2)N(\mu_c,2I_2).

The means are:

  • BA-Shapes, classes 0–3: (0,1), (1,0), (−1,0), (0,−1).
  • Tree-Grid: (0,1) for class 0 and (−1,0) for class 1.
  • BA-2Motifs: (0,1) for background nodes, (1,1) for cycle-motif nodes and (−1,−1) for house-motif nodes.

Your question helped us find a bug in the BA-2Motifs generator: background-node distributions depended on the graph class and could therefore provide an unintended prediction signal. We corrected the generator, retrained the backbones, and updated the graph-classification results. Background nodes now have the same distribution in both graph classes.

Q4. Explanation size and conversion of edge masks. All explainers are evaluated at the same explanation size. Each output is converted to a ranked node list and truncated to the top ll nodes with l=5l=5 for BA-Shapes and BA-2Motifs, l=9l=9 for Tree-Grid, and l=3l=3 for MUTAG.

For edge-mask methods, a node’s score is the average weight of its incident edges. Baselines use their default hyperparameters. We have added these details to the revised appendix.

We also evaluated different explanation sizes on BA-Shapes:

Explanation size (l)IoUPrecisionRecall
330.58890.58890.98610.98610.59170.5917
440.76910.76910.97400.97400.77920.7792
550.95260.95260.97100.97100.97100.9710
660.82270.82270.82640.82640.99170.9917
770.71050.71050.71150.71150.99610.9961

Larger explanations recover more motif nodes while including more non-motif nodes. The explanation size still has to be chosen when the motif size is unknown.

During these checks, we found that the GMM sometimes failed to converge. Increasing the number of initializations and maximum iterations resolved the observed failures. We updated the affected BA-Shapes results.

Q5. Identical and incorrectly copied results. PGExplainer’s Tree-Grid/GraphSAGE precision had been copied from the IoU column. We reran the experiment ten times, obtaining IoU 0.1718 ± 0.0386 and precision 0.2659 ± 0.0471, with sample standard deviations.

The identical InputXGradient and GNNExplainer summaries on Tree-Grid/GIN were checked and retained. Both methods consistently failed in that setting.

We have also removed the statement that predictive accuracy bounds explanation IoU. Classification accuracy and motif recovery measure different properties, so their numerical values need not follow that ordering.

AuthorsOpenReview ↗

2. Textual evaluation: metric scope, human scores, and corrected comparisons

Node Overclaiming and Structure. We agree that “Node Fidelity” was a misleading name and have renamed it Node Overclaiming. It checks whether the explanation assigns a selected node to motif and cycles when it does not belong to it. Higher scores mean less overclaiming.

CIExplainer’s set of selected nodes can include nodes outside the ground-truth motif (e.g., node 1 in Fig. 2a), and a verbalizer asked to explain “why these nodes support the prediction” may attribute a structural role for them. Because BA-2Motif fixes the motif nodes, any claim that such a node belongs to a motif is verifiably false, so ground truth is the right reference here. The criterion deliberately does not penalise silence or uncertainty about noise nodes, and does not penalise ground-truth nodes the explainer missed. The text may say a node is influential for the model without placing it in the motif.

The Structure criterion checks whether the described connections and motifs agree with the supplied graph facts. These criteria evaluate the text’s factual accuracy.

On the pairwise win-rate. We agree the win-rate should not be read as independent evidence of explanation quality, and we have toned down its role in the text. We now clearly state what the pairwise judge is evaluating (an holistic evaluation of structure, clarity, coverage and responsiveness). Since submission, we have added a rule-based baseline and a new prompt (P5) asks the LLM to verbalize facts produced by a deterministic template. We re-ran the pairwise comparison over all methods (updated Table 3), so the figures differ from the 89% in the submitted version. The new method (P5) is favored almost always (96.5%), and P4 now obtains 58.9%, the highest among the prompts in which the LLM reasons over the graph itself, while P1–P3 remain below 50%. We note, however, that the win-rate cannot be explained by constraint compliance alone. P3 and P4 share the same reasoning rules and scope constraints and differ only in input representation (raw 1-hop edges vs. detected motifs). Their win-rates still differ substantially (35.5% vs. 58.9%), so the representation, not the instructions alone, drives the gap.

The human Factual Consistency Score combines Correct Class, Correct Nodes, Correct Structure, and No Hallucination, with the sum mapped to a 1–5 scale. P2 and P4 tie at 4.333, while P3 scores 3.095. We now describe this non-monotone pattern. P4’s human-evaluation advantage over P2 is in Clarity and Overall Quality.

Corrected tables and full-test reevaluation. We thank the reviewer for identifying the discrepancy. A table contained a transcription error, and the appendix discussion did not match the reported comparisons. The earlier full-test table favored P4 on Node Fidelity and Structure, and GraphXAIN on Clarity. We have corrected the inconsistency and removed the unsupported length-bias explanation.

In the reevaluation, Structure is assessed with Claude Sonnet 5.5, while Node Overclaiming and Clarity are assessed with Qwen 3.8 27B.

On the 20-graph subset and on the full test set, P5 exceeds GraphXAIN on all three criteria.

The updated results for the full test set are:

MethodNode Overclaiming ↑Structure ↑Clarity ↑
GraphXAIN adapted1.951.122.29
P4 Motif-Aware + R.4.5433.2382.979
P5 Rule-based + reasoning5.004.064.23
Reviewer_4apeOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CIExplainer++, a two-stage framework for explaining Graph Neural Network (GNN) predictions. CIExplainer identifies important nodes through feature-level interventions, while G2TeXplainer generates natural language explanations using motif-aware prompting.

The framework is evaluated across four GNN architectures on node and graph classification benchmarks, with additional LLM-based and human evaluations of textual explanations on BA-2Motifs.

Critical strengths
  1. The framework is straightforward and practically meaningful, combining perturbation-based subgraph explanations with natural language descriptions to improve the accessibility of GNN predictions.

  2. The evaluation covers multiple GNN architectures and established baselines. CIExplainer achieves promising results on synthetic node classification benchmarks, supported by ablation studies on perturbation strategies.

Critical weaknesses
  1. Limited robustness and methodological novelty. CIExplainer substantially underperforms established methods in several graph classification settings. For example, its IoU is 0.1349 versus 0.7696 for PGExplainer on BA-2Motifs with GAT, and 0.1160 versus 0.6251 for GNNExplainer on the real-world molecular dataset with GCN. These gaps raise concerns about its generalizability. Furthermore, the methodological contribution relative to the earlier CIExplainer work (Kumar et al., 2024) is insufficiently clarified.

  2. Insufficient validation of the complete framework. The two-stage pipeline is evaluated only on BA-2Motifs, a synthetic dataset with predefined motifs. The P4 prompt explicitly provides the predicted label, selected nodes, and detected structures, making it unclear how much the LLM contributes beyond verbalizing structured information. Moreover, the framework lacks quantitative evaluation of whether it can faithfully identify insufficient or misleading evidence. These limitations weaken the claims of interpretability and faithfulness.

Questions for authors
  1. What are the main methodological differences between the proposed CIExplainer and Kumar et al. (2024)?

  2. Why does CIExplainer underperform established baselines in several graph classification settings? Are explanation-size budgets and evaluation protocols comparable?

  3. Why is the complete pipeline evaluated only on BA-2Motifs? Are textual explanation results available for node classification or real-world graphs?

  4. How are the motifs supplied to G2TeXplainer extracted? Do they rely on predefined rules or ground-truth annotations? Could the authors compare P4 against a rule-based template using the same structural inputs?

  5. Although P4 explicitly instructs the model to recognize unsupported evidence, how reliably does it generate negative explanations when selected nodes do not support the prediction?

5 comments
AuthorsOpenReview ↗

Thank you for your review

We thank the reviewer for the assessment and questions. Your comments helped us identify a BA-2Motifs feature-generation bug and a GMM convergence issue, and improve the evaluation. We corrected the affected results, clarified the relationship to Kumar et al. (2024), added a rule-based comparison and P5, and audited negative cases for P5. CIExplainer remains the main contribution, with G2TeXplainer as a secondary component. Changes are marked in blue in the revised manuscript. The following subcomments address all five questions.

AuthorsOpenReview ↗

1. CIExplainer: contribution, graph-classification results and evaluation protocol (Q1–Q2)

Relationship to Kumar et al. (2024). The earlier work is a nonarchival, two-page workshop extended abstract describing preliminary CIExplainer for bipartite link prediction, illustrated through a medical-referral example. The present submission retains its feature-intervention and node-ranking principle. It develops an explicit finite-budget intervention-effect estimand and distribution-based counterfactual generation, including Gaussian-mixture sampling, extending the earlier rule based on distant observed values. We also clarify the intervention assumptions. We have made the extension of Kumar et al. (2024) explicit in Appendix A.4 (p. 17).

The evaluation covers BA-Shapes, Tree-Grid, BA-2Motifs and MUTAG across GCN, GAT, GIN and GraphSAGE, with ablations of aggregation, sampling budget, intervention distributions and marginal versus joint sampling. G2TeXplainer adds textual descriptions, evaluated using LLM-as-judge and human evaluation.

Graph-classification results. The original BA-2Motifs feature generator used graph-class-dependent distributions even for background nodes, creating a prediction shortcut. The correction gives background nodes a common distribution across graph classes. We retrained all backbones and reran the graph-classification comparisons. CIExplainer now ranks first on three of four BA-2Motifs backbones; InputXGradient and Shapley Values also improve substantially. GAT remains challenging, so performance is not uniform across architectures.

On MUTAG, CIExplainer’s mean performance remains above random across all four backbones. With GIN, it outperforms GNNExplainer, PGExplainer and SubgraphX. We acknowledge the gap to GNNExplainer with GCN and the resulting limitation on generalization.

Comparable evaluation. Each explainer’s output is converted to a ranked node list and truncated to the same size (l): 5 for BA-Shapes and BA-2Motifs, 9 for Tree-Grid, and 3 for MUTAG. Edge masks are converted to node scores by averaging incident edge weights, The PGExplainer paper [1] used an equivalent approach to turn node weight methods into edge weights. Baselines use default hyperparameters. The protocol is now explicit in the revised appendix.

Because this comparison uses the ground-truth motif size, we also evaluated CIExplainer’s sensitivity to explanation size on BA-Shapes:

Explanation size (l)IoUPrecisionRecall
30.58890.98610.5917
40.76910.97400.7792
50.95260.97100.9710
60.82270.82640.9917
70.71050.71150.9961

Smaller explanations have higher precision but omit motif nodes. Increasing (l) improves recall while including more non-motif nodes. This measures budget sensitivity; selecting (l) without ground-truth information remains a limitation.

During this evaluation, we also identified GMM convergence failures. Increasing the number of initializations and maximum iterations resolved the observed failures. We updated the affected BA-Shapes results in Table 2; CIExplainer now outperforms the compared methods across all four BA-Shapes backbones.

We have added the common evaluation protocol in Appendix A.1.3 (p. 16) and the explanation-size analysis in Appendix A.6 and Table 6 (pp. 20 and 22).

[1] Luo, D., Cheng, W., Xu, D., Yu, W., Zong, B., Chen, H., & Zhang, X. (2020). Parameterized explainer for graph neural network. Advances in neural information processing systems, 33, 19620-19631.

AuthorsOpenReview ↗

2. G2TeXplainer: evaluation scope, structural inputs and negative cases (Q3–Q5)

Evaluation scope (Q3).

We chose BA-2Motifs for textual evaluation because both graph classes have known, distinct explanatory motifs: houses and cycles. This allows checking whether descriptions distinguish the structural evidence associated with alternative predictions.

Our current node-classification evaluation considers motif nodes, a setting useful for many problems where the goal is to explain an individual node’s prediction from its features and neighborhood. The repeated motif structures in these benchmarks offer limited variation for evaluating textual descriptions. On MUTAG, explanation evaluation includes only mutagenic graphs, because the benchmark provides explanatory groups for those graphs. Non-mutagenic graphs are included in GNN training but excluded from explanation evaluation because they lack ground-truth explanation subgraphs.

These restrictions make those protocols less informative for assessing whether textual descriptions distinguish evidence for different classes. We therefore used BA-2Motifs for the textual evaluation, while evaluating CIExplainer across all four datasets. We report no textual results for node classification or real-world graphs; extending that evaluation remains future work depending on the existence of labeled datasets.

We have clarified which instances are evaluated in Appendix A.1.3 (p. 16) and acknowledged the textual component’s generalization limits in Section 5 (p. 10).

Motif extraction and rule-based comparison (Q4).

We detect predefined patterns, including triangles, 4-cycles and 5-cycles, automatically from graph topology using NetworkX. Ground-truth motif annotations are not used for extraction. The prompt receives the GNN’s predicted label, selected nodes, structural definitions and detected structures.

We thank the reviewer for this suggestion to consider a rule-based template. We added two rows to Table X: (i) a Rule-base baseline, which fills a fixed template deterministically from the same detected structures; and (ii) P5 (Rule-base + R.), which verbalizes the template’s facts with reasoning instructions.

The most relevant mean judge scores are:

MethodNode overclaim score ↑Structure ↑Clarity ↑
Rule-based template5.0005.0002.762
P44.5433.2382.979
P54.9524.2383.952

The template achieves the highest structural scores. P5 improves clarity, while retaining some structural errors. This comparison supports the LLM’s contribution to readable verbalization of supplied facts.

We have added the rule-based baseline and P5 in Section 3.4 (pp. 6–7), with their comparison reported in Table 3 and discussed in Section 4.2 (pp. 8–9).

Negative cases (Q5).

For the new P5 variant, we manually audited outputs from one fixed seed against the supplied structural facts. In 18 graphs, at least one selected node lay outside the detected structures. None of the explanations assigned a structural role to these nodes: 16 explicitly acknowledged their absence from the structures, and two omitted them.

In the two graphs where no selected node belonged to the motif, both explanations concluded that the selected nodes did not justify the prediction. In mixed cases, the structural justification referred to motif nodes.

The audit also identified overstatements about model behavior. Of the 16 explanations that discussed nodes outside the detected structures, 10 confined their statements to structural membership, four said these nodes “do not support the prediction,” and two said they “do not contribute.” Absence from a detected structure does not establish absence of influence on the GNN, so those six descriptions exceeded the supplied evidence.

We have clarified the Node Overclaim criterion in Appendix A.8 (p. 25) and report the additional manual P5 audit.

Comments on the paper

1 comment

We thank the reviewers for their comments and for recognizing the paper’s contributions. We appreciate 4ape’s assessment of the framework’s practical value, its accessibility through textual descriptions, and the breadth of the backbone and baseline comparisons. KUM7 highlighted the forward-pass computation, stability across all four backbones on node-classification benchmarks, and repeated ablations isolating the effects of aggregation, sampling budget, intervention distribution, and marginal versus joint sampling. RhKs valued the potential-outcome formulation, its applicability to node classification, graph classification and link prediction, and the reported efficiency and textual-evaluation results.

The revision builds on these strengths. CIExplainer is presented as the main methodological contribution, with clarified absolute-effect aggregation and computation-graph assumptions. We corrected the BA-2Motifs feature generator, resolved observed GMM convergence failures, and reran affected experiments. We also documented common explanation sizes and edge-mask conversion.

The main changes are:

  • Corrected experimental results. We corrected a BA-2Motifs feature-generation bug that introduced a graph-class signal into background nodes, retrained the backbones, and reran the comparisons. We also resolved observed GMM convergence failures affecting BA-Shapes. In the updated comparisons, CIExplainer achieves the highest mean IoU on all four BA-Shapes backbones and three of four BA-2Motifs backbones. Performance on MUTAG remains architecture dependent.
  • Clarified formulation and evaluation. We corrected the definitions to specify maximum absolute effects, documented common explanation sizes and edge-mask conversion, and corrected transcription errors. Our response to RhKs specifies the explanation instance, fixed inputs, and the conditions that support the computation-graph restriction. The causal interpretation concerns interventions on the trained GNN.
  • Revised textual evaluation. G2TeXplainer is presented as a secondary component and LLM-as-judge as an evaluation procedure. We added a deterministic template baseline and P5, corrected inconsistent comparisons, and distinguished textual factual accuracy from model faithfulness. The manual audit also reports residual overstatements about model behaviour.

We ask that the assessment consider the corrected results and clarified scope alongside the practical value, cross-backbone stability, and computational approach,

Thank you, all the reviewers, for your thoughtful comments! Please let us know if there remain unresolved concerns.

The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.