acceptodds
Under review as a conference paper at ICLR 2027

STAR: A Transferable Explanatory Approach for Graph Neural Networks

Abstract

Graph neural networks (GNNs) integrate node attributes and relational structure through neighborhood aggregation to capture graph dependencies, but nonlinear computations obscure the decision evidence underlying individual predictions. To improve explainability, post-hoc explanatory methods identify decision evidence without modifying the pretrained GNN's architecture or parameters. Among these approaches, parameterized explanatory methods exploit cross-instance explanatory information to improve explanations for a given pretrained GNN. Extending these explanatory methods to other pretrained GNNs requires accounting for differences in predictive behavior. Although previously acquired explanatory information can guide explanation generation, directly applying the fitted explanation function risks misidentifying the new pretrained GNN's decision evidence. Refitting can require additional samples that are unavailable at explanation time. To address this challenge, this paper proposes STAR, a transferable explanatory approach for GNNs based on Source-informed explanatory prior generation and Target-Adaptive explanation Refinement. To exploit explanatory information across pretrained GNNs, STAR develops a response-conditioned explanatory prior generator that models explanatory relevance in relation to the queried pretrained GNN's behavior and input information. Prediction-aligned optimization across source pretrained GNNs enables explanatory guidance for new pretrained GNNs without generator refitting. To preserve explanation fidelity despite potential source-to-target mismatches, STAR develops target-adaptive explanation refinement that evaluates evidence guided by the explanatory prior alongside locally identified alternatives under the target pretrained GNN. This assessment allows useful explanatory prior guidance to inform the final explanation while permitting correction through target-specific evidence. Experiments across six datasets and three GNN architectures demonstrate competitive explanatory fidelity and effective transfer to target pretrained GNNs excluded from generator fitting.

Primary area:
interpretability and explainable AI
Keywords:
Graph Neural Networks, Post-hoc Explainability, Transferable Explanations, Explanatory Priors, Subgraph Explanations

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_HgboOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes STAR, a transferable approach for explaining the predictions of pretrained graph neural networks (GNNs). STAR first trains a generator using multiple source GNNs to produce initial edge importance scores based on the input structure and the representations and predictive responses of the model being explained. It then performs local optimization, subgraph search, and candidate selection on the target model to refine potentially unsuitable source-derived guidance. Both the generator and the target GNN remain frozen during explanation. The authors conduct comparative and ablation experiments across six datasets and three GNN architectures. The results show that the complete procedure achieves competitive explanation fidelity in most evaluated settings.

Critical strengths
  1. The paper addresses a practically meaningful problem in explanation transfer: Many parameterized explainers are trained for a particular model, and their effectiveness may not be preserved when the model being explained changes. Refitting an explainer may also require additional target samples.

  2. The experimental evaluation is extensive. The paper includes multiple datasets, several model architectures, component ablations, and cross-domain tests. These experiments examine the complete method from different perspectives and provide empirical support for its effectiveness and transferability within the evaluated settings.

Critical weaknesses

STAR uses pretrained source GNNs to train a generator that produces initial explanation scores. However, whether the knowledge learned by the generator from these source models is representative and applicable to the target model’s decision-making process requires further validation. Although the paper includes held-out and cross-domain evaluations, it does not sufficiently establish how source model quality, diversity, and differences from the target model affect the initial scores and final explanations. Consequently, it remains unclear how much benefit the generator provides when available source models are limited or differ substantially from the target model, and whether subsequent refinement can adequately compensate for unsuitable guidance.

Questions for authors

How sensitive is STAR to the choice of source models? When source models are of lower quality, fewer in number, or rely on different decision cues from the target model, how do the quality of the initial edge scores and the final explanations change? To what extent can target-specific refinement compensate for unsuitable source-derived guidance?

1 comment

Thank you for your valuable question. We examine the initial and final explanations under limited source availability and a held-out target architecture, and explain how STAR handles potentially unsuitable source guidance.

Firstly, in each run, the explanatory prior generator is fitted using only three source GNNs pretrained on MUTAG, one GCN, one GAT, and one GIN. It then explains target GNNs pretrained on BBBP. BBBP is excluded from generator fitting, and search width is fixed without selection on BBBP. Results are averaged over three seeds.

Target combinationGenerator-only Fidelity AUCSTAR Fidelity AUCGenerator-only Robust Fidelity AUCSTAR Robust Fidelity AUC
BBBP–GCN0.73160.84880.72150.8408
BBBP–GAT0.76250.85780.76110.8560
BBBP–GIN0.75800.81540.77090.8270
Average0.75070.84060.75110.8413

With this restricted source collection from a different dataset, refinement improves average Fidelity AUC by 0.0899 and Robust Fidelity AUC by 0.0902. These comparisons quantify the improvement over explanations derived from the initial edge scores when only a small source collection is available.

Secondly, we conduct an additional leave-one-architecture-out experiment. Every GCN–dataset combination is excluded from generator fitting and inference-configuration selection. Only GAT and GIN models across the six datasets participate in these procedures. The held-out GCNs are used exclusively for target evaluation, without generator refitting. Results are averaged over three seeds.

Held-out target combinationGenerator-only Fidelity AUCSTAR Fidelity AUCGenerator-only Robust Fidelity AUCSTAR Robust Fidelity AUC
MUTAG–GCN0.61330.63930.60540.6324
BBBP–GCN0.74760.84970.74130.8404
BACE–GCN0.56360.57430.56190.5722
HIV–GCN0.68800.71360.69620.7186
Benzene–GCN0.63510.78230.58990.7580
PROTEINS–GCN0.61240.65640.61170.6554
Average0.64330.70260.63440.6961

Refinement improves both dataset-level mean metrics on all six datasets despite the absence of the target architecture from the source collection. Average Fidelity AUC increases by 0.0593, and Robust Fidelity AUC increases by 0.0617.

Regarding lower-quality source models or different decision cues, their influence enters STAR through the explanatory prior learned by the generator. At explanation time, the generator also conditions on the target GNN’s input information and predictive responses, rather than simply reproducing a source-associated explanation. Refinement then assesses prior-guided candidates alongside locally optimized alternatives under the target GNN. Source-derived guidance therefore influences the search without uniquely determining the final explanation.

This design addresses source–target differences by using source-acquired information to guide candidate construction and the target GNN’s predictive responses to assess the resulting explanations. Locally optimized alternatives allow refinement to select evidence beyond the generated prior when that guidance is unsuitable. The improvements in both tables demonstrate the benefit of this target-specific refinement under limited source coverage and an unseen target architecture.

Reviewer_6qynOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies the transferability of parameterized GNN explainers across pretrained models. The setting assumes access to the target pretrained GNN and the instance to be explained, without additional target samples for refitting the explainer. The authors argue that directly transferring a source-fitted explainer can produce a mismatch with the evidence supporting the target GNN’s prediction. To address this issue, they propose STAR, which first learns a response-conditioned explanatory prior generator from graph structure, representation descriptors, and predictive responses. At target time, the generated edge-relevance scores guide candidate construction and search, while target-adaptive explanation refinement determines the final explanatory subgraph. Experiments cover six datasets and three GNN architectures, reporting Fidelity AUC and Robust Fidelity AUC.

Critical strengths

Developing explainers that maintain their fidelity when transferred across GNN architectures is a valuable research direction.The complete STAR method achieves higher Fidelity AUC and Robust Fidelity AUC than the evaluated baselines in most dataset–architecture configurations.The appendix provides detailed experimental settings and several useful controls for assessing the method.

Critical weaknesses

The existence of the motivating problem is not empirically established: The paper is motivated by a mismatch between source-based explanations and the evidence supporting target-model predictions. However, Figure 1 is only a schematic illustration and does not explain how the depicted “target-supported evidence” is established. If this evidence is identified through prediction preservation, multiple subgraphs may satisfy that criterion, and the preference for the amide-containing fragment requires model-based justification. If it is based on chemical knowledge, chemical plausibility alone does not establish faithfulness to the target GNN. The paper does not directly quantify the additional fidelity loss incurred when an existing parameterized explainer is transferred between independently trained GNNs on the same task. Consequently, the severity of the motivating problem and the extent to which STAR resolves it remain unclear.

The contributions and rationale of STAR’s two-stage design are insufficiently established: The generator uses the queried GNN’s representations and predictive responses to provide model-dependent explanatory guidance, while target-adaptive refinement further optimizes and selects explanatory subgraphs. However, it remains unclear to what extent each stage addresses the source-to-target mismatch and what additional benefit is obtained by combining them.

Under Joint-fit, the architecture-averaged Fidelity AUC on BBBP is 0.8402 for Local-only and 0.8443 for complete STAR; the corresponding values on MUTAG are 0.6342 and 0.6414. Although complete STAR improves over Local-only in these settings, the incremental gains in the reported means are modest and do not clearly establish the learned prior’s contribution. Overall, the current results do not sufficiently substantiate the benefits of the two-stage design, and the paper does not clearly explain the rationale for combining the explanatory prior generator with target-adaptive refinement.

A gap between the motivation and the experimental comparisons: The motivating claim concerns the failure of a source-fitted parameterized explainer when directly applied to another GNN. A directly corresponding evaluation would compare source-to-target transfer against refitting the same explainer on the target GNN. The main baseline comparisons do not measure this transfer penalty. They evaluate the quality of STAR’s final explanations, but do not establish how severe the original mismatch is or how much of that loss STAR recovers.

Limited statistical evidence and robustness validation: The critical difference analysis in Figure 3 is based on 108 interdependent evaluation tasks, for which the independent-block assumption is not established. The authors also explicitly characterize this analysis as an exploratory reference.

Questions for authors

For independently trained source and target GNNs with the same architecture and task, how much fidelity is lost when a source-fitted parameterized explainer is directly transferred, relative to refitting that explainer on the target GNN? How much of this loss does STAR recover?

How do the authors explain the near-complete-method performance of some variants without prediction alignment, and the relatively small gap between Local-only and complete STAR?

1 comment

Response to Weaknesses 1 and 4 and Question 1:Thank you for this valuable question. To substantiate the motivation, we use ReFine as an example and examine transfer between different dataset–GNN combinations. Within each comparison, all explanatory methods explain the same target GNN on identical test instances. We report Fidelity AUC averaged over three seeds.

Source combinationTarget combinationReFine direct transferReFine target refittingSTAR
BACE–GCNBBBP–GCN0.62420.79470.8505
BACE–GINBBBP–GIN0.69010.76830.8176
BACE–GATBBBP–GAT0.74980.78220.8621
BBBP–GATBACE–GAT0.51690.55560.5784

In these examples, directly transferred ReFine loses 0.1705, 0.0782, 0.0324, and 0.0387 Fidelity AUC relative to target-refitted ReFine. These results demonstrate that directly applying source-acquired explanatory information to a target GNN can reduce explanatory fidelity, providing empirical support for the motivation. STAR achieves higher fidelity in these settings without target-specific generator refitting. The source column specifies ReFine’s fitting source. STAR uses multiple source GNNs, with the corresponding target dataset excluded from generator fitting.

For the same-task, same-architecture setting specifically requested, consider two independently fitted GINs on BBBP. ReFine fitted to one GIN and directly applied to the other achieves 0.7151, compared with 0.7635 when fitted to the target GIN, a loss of 0.0484. STAR achieves 0.8176 on the same target instances, exceeding these conditions by 0.1025 and 0.0541, respectively. Thus, explanatory mismatch can also arise between independently pretrained GNNs with identical tasks and architectures.

Figure 1 provides a schematic example to help readers understand this mismatch. The controlled comparisons above provide the empirical evidence for the motivation.

Response to Weaknesses 2 and 3 and Question 2:Thank you for examining the component contributions. The following table reports joint-fit Fidelity AUC averaged over GCN, GAT, and GIN and three seeds.

DatasetGenerator-onlyLocal-onlyWithout prediction alignmentSTAR
MUTAG0.62650.63420.63470.6414
BBBP0.77770.84020.84120.8443
Benzene0.68930.70170.70420.7209

The small gap on BBBP reflects the strong explanatory performance already achieved by target-side optimization. Local-only reaches 0.8402, whereas Generator-only reaches 0.7777. STAR retains locally optimized candidates alongside prior-guided candidates. When local optimization already identifies evidence with strong predictive support, the prior has less room to improve the final explanation. Its contribution is consequently incremental rather than additive. On Benzene, the improvement over Local-only is larger, from 0.7017 to 0.7209, showing that the value of prior guidance depends on the target setting.

The same distinction helps explain why removing prediction alignment can leave final performance close to STAR. This variant retains target-side optimization and candidate assessment, which can compensate for less useful generator guidance. Final fidelity therefore need not change in proportion to changes in the generated prior. This is a mechanism-based explanation, rather than evidence that compensation occurs in every case. The observed alignment gains are 0.0067 on MUTAG, 0.0031 on BBBP, and 0.0167 on Benzene.

The rationale for combining the components is therefore to make source-acquired guidance available without requiring it to determine every explanation. Target-side refinement remains effective when local evidence is sufficient, while prior-guided candidates provide additional opportunities for improvement. STAR’s highest average fidelity across these three comparisons supports this complementary design.

Response to Weakness 5:Thank you for identifying the dependence among evaluation entries. We will present the CD plot as a descriptive summary of relative rankings across settings and will not use its significance thresholds to support claims of statistical superiority.

Reviewer_q73HOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper presents a method for transferring explanatory information through pretrained GNNs using a response conditioned prior generator and target specific refinement. The method uses a frozen generator and uses local mask optimisation and subgraph search to identify evidence to support the target prediction. The paper performs several experiments to support it's claims, including evaluation of explanatory fidelity on 6 datasets and 3 GNNs, and a transfer study on Cora dataset.

Critical strengths

The problem is important and useful for the community.

  1. Fitting a new explainer for every trained GNN is expensive, and reusing explanatory information could reduce that need.
  2. Presented experiments have useful controls for model conditioning, generator free inference, and consistency of explanations. The full method is generally achieving strong results than the evaluated baselines.
  3. Appendix is transparent about limitations.
Critical weaknesses

My main concerns are of novelty, presented baselines and evaluation setup:

  1. Can the authors clearly present novelty and comparison against ReFine and TAGE? Both are discussed in text, but not included in experimental comparison. Specifically ReFine also uses learned explanatory patterns with adaption per instance. TAGE does have a different setting with a fixed embedding model, but a fair adaptation would allow stronger arguments.
  2. Search based methods like SubgraphX, EiG search, etc. are also not compared - thus it is hard to asses STAR's contribution beyond the local search. Can some baselines be provided?
  3. The ablation of transfered information is not (sufficiently) separated from additional search. Since STAR is using ~2.3x many forward calls as local only, and against uniform/random initialisation don't shoe clear gains. For example - on BBBP, fidelity improves by ~0.01 for STAR but since it also performs discrete search at same retained fractions used for evals, comparison with matched search and query budget would be needed to the contribution of learned prior. Can authors explain this?
  4. The held out setting seems weak to me (than the problem motivation suggests). Evaluated dataset are excluded from generator fitting, but can still contribute largely to hyperparam selection.
  5. Robust fidelity explains each perturbed input using its own predicted class, so it would measure predictive support after re-explanation than the stability of the original explanation.
  6. Ablations are not very supportive. For example its not clear to me the benefit of the perturbation conditioned training. In table A.9, generator only fidelity lowers in 4/6 comparison. While complete method is performing well, the ablations are not supporting the contribution of the perturbation objectives and selection rules.
Questions for authors

See weaknesses for questions

4 comments
AuthorsOpenReview ↗

Rebuttal Part I

Response to Weakness 1: Thank you for your valuable question. ReFine learns explanatory patterns and adapts them to individual instances. STAR specifically conditions its explanatory prior on the queried pretrained GNN and compares prior-guided candidates with locally optimized alternatives. Its distinction is how source-acquired information is adapted to a different pretrained GNN, rather than instance-specific refinement alone.

TAGE transfers explanations across downstream tasks associated with a GNN embedding model. STAR instead addresses changes in the pretrained GNN being explained. Comparing TAGE under this setting requires extending its original transfer protocol. We therefore prioritize ReFine as the closer comparison involving learned explanatory information and instance-specific refinement.

ReFine is evaluated across six datasets, three architectures, and three seeds, including its instance-specific refinement. Selected seed-averaged Fidelity AUC results are below.

TargetReFineSTAR
MUTAG–GIN0.53290.5984
BBBP–GIN0.75260.8188
PROTEINS–GIN0.59380.6108

Across the complete evaluation, STAR leads on both Fidelity AUC and Robust Fidelity AUC in 16 of 18 dataset–architecture combinations.

Response to Weakness 2: Thank you for this suggestion. We conduct a search-based comparison and prioritize a SubgraphX example because of its substantial evaluation cost. This completed comparison uses BBBP–GIN, seed 0, identical target instances, and the shared edge-ranking AUC protocol.

MethodFidelity AUCRobust Fidelity AUC
SubgraphX, 10 rollouts0.78700.8016
SubgraphX, 20 rollouts0.79220.8024
STAR held-out0.83320.8476

Increasing SubgraphX’s rollout count improves its Fidelity AUC, while STAR retains higher performance on both metrics in this example.

Response to Weakness 3: The reported forward calls cover the entire explanation procedure, including target-response collection, local optimization, candidate search, and intervention-based assessment. STAR evaluates both prior-guided and local candidates. The 2.3-fold query count therefore does not mean 2.3 times as many local optimization iterations.

The additional candidate route nevertheless consumes target-GNN queries. The original component ablations compare the complete prior-guided procedure with local inference, rather than isolating learned information at an identical total query budget. Uniform/random controls examine the prior’s information content when candidate construction and search settings are held fixed. We will distinguish these comparisons explicitly. Matching retained fractions controls explanation size, not computational expenditure.

Response to Weakness 6: Thank you for this question. The perturbation-conditioned objective encourages the explanatory prior to retain predictive support under input changes, emphasizing cases where the pretrained GNN’s prediction changes little. We assess its contribution by removing the intervention-dependent penalty while retaining the same refinement and candidate-selection procedure.

For example, on MUTAG–GIN under held-out fitting, the final results averaged over three seeds are as follows.

Training configurationFidelity AUCRobust Fidelity AUC
Without the intervention-dependent penalty0.59050.5946
Complete objective0.59580.5961

Regarding the four decreases in Generator-only Fidelity in Table A.9, these results indicate that the objective does not uniformly improve explanations extracted directly from the generated scores. Those scores also guide candidate construction and search, so their standalone fidelity does not fully describe their contribution to the final explanation. The comparison above evaluates that contribution after refinement and shows improvements in both metrics in this setting.

Candidate selection assesses prior-guided and locally optimized explanations using the target GNN’s predictive responses. This allows source-acquired guidance to inform the explanation while enabling target-specific evidence to replace unsuitable guidance. Because this step determines which candidate becomes the final explanation, simply deleting it leaves the output undefined and therefore does not provide a meaningful ablation of its contribution.

AuthorsOpenReview ↗

Rebuttal Part II

Response to Weakness 4: Thank you for this question. We conduct two strict holdout experiments, excluding the held-out combinations from both generator fitting and inference-configuration selection. All comparisons use the same frozen target GNNs and test instances. Results for the six original baselines are reused from the main experiments. The additional ReFine comparison includes target-specific fitting and instance-specific refinement.

Each entry reports Fidelity AUC / Robust Fidelity AUC, averaged over three seeds.

First, every BBBP–architecture combination is excluded. Generator fitting and configuration selection use only the other five datasets.

TargetGNNExplainerPGExplainerGraphMaskGEARIDEAGraph CurvaturesReFineSTAR strict holdout
BBBP–GCN0.7667 / 0.77070.6609 / 0.66560.6878 / 0.66870.7348 / 0.72920.6743 / 0.67200.6920 / 0.69170.7796 / 0.77630.8494 / 0.8403
BBBP–GAT0.8015 / 0.80260.7358 / 0.73260.7397 / 0.74430.7777 / 0.77810.7574 / 0.75450.7425 / 0.74460.7853 / 0.78450.8716 / 0.8671
BBBP–GIN0.7466 / 0.76390.6688 / 0.68040.7040 / 0.71890.7873 / 0.79440.5941 / 0.61110.6689 / 0.68610.7526 / 0.75810.8185 / 0.8295

STAR achieves higher values on both metrics for all three architectures without BBBP participating in generator fitting or configuration selection.

Second, every dataset–GCN combination is excluded. Generator fitting and configuration selection use only GAT and GIN models across the six datasets. All GCNs are reserved for target evaluation.

TargetGNNExplainerPGExplainerGraphMaskGEARIDEAGraph CurvaturesReFineSTAR strict holdout
MUTAG–GCN0.6113 / 0.60750.6124 / 0.60260.6057 / 0.59970.6090 / 0.60260.6093 / 0.60190.6167 / 0.60640.6181 / 0.61280.6393 / 0.6324
BBBP–GCN0.7667 / 0.77070.6609 / 0.66560.6878 / 0.66870.7348 / 0.72920.6743 / 0.67200.6920 / 0.69170.7796 / 0.77630.8497 / 0.8404
BACE–GCN0.5625 / 0.56110.5492 / 0.54830.5542 / 0.55450.5601 / 0.55840.5489 / 0.54860.5600 / 0.55800.5649 / 0.56280.5743 / 0.5722
HIV–GCN0.6622 / 0.67730.5952 / 0.62160.6076 / 0.63800.6604 / 0.67480.6011 / 0.62850.6114 / 0.63520.7166 / 0.72050.7136 / 0.7186
Benzene–GCN0.6155 / 0.59830.6599 / 0.61970.6619 / 0.60180.6538 / 0.61130.6582 / 0.61460.6599 / 0.61980.6674 / 0.64590.7823 / 0.7580
PROTEINS–GCN0.6277 / 0.62710.5755 / 0.57570.5588 / 0.59990.5967 / 0.59650.5690 / 0.56940.5803 / 0.58050.6224 / 0.62300.6564 / 0.6554

STAR leads on both metrics in five of the six held-out GCN combinations. On HIV–GCN, ReFine performs slightly better. ReFine is fitted using training instances and predictive responses from the target HIV–GCN, whereas STAR’s generator fitting and configuration selection exclude every GCN model. STAR therefore achieves comparable performance in this case without target-specific generator refitting. In the complete main comparison reported in our response to Weakness 1, STAR outperforms ReFine on both metrics in 16 of 18 dataset–architecture combinations.

Response to Weakness 5: Thank you for clarifying this distinction. The manuscript’s metric definition evaluates predictive support after re-explaining each perturbed input, including prediction-changing views. It explicitly distinguishes this evaluation from fixed-evidence and edge-overlap stability.

Robust Fidelity assesses whether the explanatory method identifies evidence supporting the pretrained GNN’s prediction under input changes. When the predicted class changes, the explanation is evaluated against that new prediction. Thus, it complements original-input Fidelity by assessing explanatory performance under perturbations, not preservation of the original explanation.

We will revise stability-related wording that could suggest otherwise, while retaining the separate fixed-evidence and edge-overlap diagnostics.

I have reviewed the response in detail and thank the authors for the additional experiments. Although some questions are answered - the main concerns still remain:

  • The contribution of the learned prior is still not separated from the extra search, also no query-matched comparison is provided, and the appendix says the current controls do not isolate it. Also, across the new experiments the final star numbers stay almost the same no matter how the generator was fitted, while generator-only fidelity drops clearly. To me this suggests the gains mostly come from target-side refinement and search, not from transferred knowledge
  • The baseline comparisons are still fairly limited
  • The ablations are not convincing to me, like the perturbation-objective result is a single setting with a very small gain

I also could not find a revised version of the paper reflecting these changes. I will maintain my score.

Thank you for your follow-up.

For your concern 1, similar final performance under different source-fitting configurations does not establish that the explanatory prior is unused. Generator-only measures explanations directly extracted from generated scores, while STAR also uses these scores to construct blended candidates and guide search proposals. Lower standalone fidelity can therefore coexist with useful search guidance. The final results are consistent with refinement accommodating differences in source-derived guidance. They support the effectiveness of the combined procedure, while the independent contribution of the prior requires distinguishing its information from the additional candidate evaluations.

The direct-transfer comparison provides complementary evidence. On BBBP–GIN, ReFine fitted on BACE achieves 0.6901 Fidelity AUC, compared with 0.7683 for target-fitted ReFine. STAR achieves 0.8176 with BBBP excluded from generator fitting, demonstrating effective explanation transfer without target-specific generator refitting.

For your concern 2, our baseline selection covers representative approaches to GNN explanation, including instance-specific optimization with GNNExplainer, parameterized explanation generation with PGExplainer and GraphMask, learned explanatory information with instance-specific refinement through ReFine, and subgraph search with SubgraphX. The main experiments also include recent related methods such as GEAR, IDEA, and Graph Curvatures. Together, these baselines provide broad coverage of established approaches and recent developments, including the methods most closely related to STAR’s explanatory prior and target-adaptive refinement.

For your concern 3, the perturbation-conditioned objective shapes the scores used for both explanation extraction and downstream candidate generation. On MUTAG–GIN under held-out fitting, the intervention-dependent penalty improves final Fidelity AUC from 0.59053 to 0.59581 and Robust Fidelity AUC from 0.59458 to 0.59609. This is a modest, configuration-specific improvement. The Generator-only decreases in Table A.9 concern direct extraction from the scores, whereas the final results also reflect their use during refinement. Our evidence supports evaluating these components jointly, rather than claiming that the perturbation objective independently improves every configuration.

Regarding the revised manuscript, the additional results and clarifications are provided directly in our rebuttal comments. In rebuttal period, the revised manuscript can not be uploaded.

Reviewer_q8sTOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes STAR, a transferable post-hoc explanation framework for graph neural networks (GNNs), aiming to leverage explanatory knowledge acquired from source pretrained GNNs to explain previously unseen target GNNs without refitting the explanation generator. STAR consists of two main components: source-informed explanatory prior generation and target-adaptive explanation refinement. The former employs a response-conditioned generator trained across source GNNs through prediction-aligned objectives, while the latter combines generated priors with instance-specific optimization, candidate subgraph search, and target-dependent selection. Experiments on six graph classification datasets and three GNN architectures demonstrate competitive explanatory fidelity under both joint-fitting and held-out transfer settings.

Critical strengths
  1. The paper addresses an important and relatively underexplored problem of transferring explanatory knowledge across independently pretrained GNNs without target-specific generator refitting.

  2. The proposed framework combines response-conditioned explanatory priors with target-adaptive refinement, providing a coherent approach to mitigating source-to-target explanatory mismatches.

  3. The experiments cover multiple datasets and GNN architectures, demonstrating competitive fidelity and encouraging performance under held-out generator fitting.

Critical weaknesses
  1. The held-out evaluation does not fully establish target-independent transferability. Although target datasets are excluded from generator fitting, Section 4.1 and Appendix D.3 indicate that inference hyperparameters are selected using validation configurations that include the evaluated datasets. Moreover, all target architecture families remain represented during source fitting. Consequently, the experiments primarily demonstrate transfer within familiar architecture families rather than generalization to entirely unseen GNN architectures. Stricter leave-one-dataset-out and leave-one-architecture-out evaluations would strengthen the transferability claims.

  2. The contribution of transferable explanatory knowledge is not sufficiently isolated from target-specific optimization. Although the method introduces a source-trained explanatory prior, the complete framework also performs extensive instance-specific refinement and candidate search. As shown in Table 2, the improvements over Local-only are relatively modest in several settings, while the generator-free controls in Appendix E.4 do not fully isolate the prior's contribution under matched computational budgets. Furthermore, the Cora evaluation lacks a corresponding Strict Local-only comparison. Computationally matched controls and direct source-to-target transfer baselines would better establish the value of the learned prior.

  3. The computational overhead is substantial relative to the incremental performance gains. The proposed refinement involves multiple optimization stages, candidate evaluations, and target-GNN queries. Appendix D.5 reports approximately 813 forward calls per explanation for STAR, compared with 348 for Strict Local-only, representing a 2.34-fold increase. However, the corresponding fidelity improvements remain relatively small on several datasets. The absence of strictly query-matched comparisons and reliable inference-time measurements makes it difficult to assess the practical efficiency of the framework. A systematic fidelity–cost analysis would strengthen the empirical evaluation.

  4. Target-adaptive refinement does not consistently improve explanation fidelity. While Section 3.3 motivates refinement as a mechanism for correcting source-to-target mismatches, the experimental results reveal substantial degradation in certain configurations. For example, Appendix E.9 reports that, on Benzene–GIN, refinement reduces Fidelity AUC from 0.7213 to 0.6223 under Joint-fit and from 0.7054 to 0.5998 under Held-out fitting. These results suggest a potential mismatch between the composite refinement objective and the final evaluation metric. Further failure analysis and mechanisms to prevent substantial deterioration would strengthen the reliability of the proposed refinement.

  5. The evaluation provides limited evidence regarding explanation reliability and broader generalizability. The main results emphasize confidence-based fidelity, which does not necessarily establish agreement with meaningful explanatory evidence or explanation stability. Moreover, the supplementary analyses reveal inconsistent benefits from certain optimization components and identify several target GNN checkpoints with single-class predictions. Although these limitations are partially acknowledged, their combined effects raise questions about the robustness and generality of the reported improvements. More direct evidence-level evaluation and controlled experiments on reliable target predictors would strengthen the conclusions.

Questions for authors

Please see the weaknesses.

2 comments
AuthorsOpenReview ↗

Rebuttal Part I

Response to Weakness 1: Thank you for your suggestion. We evaluate two stricter holdout protocols that exclude the target group from both generator fitting and inference-hyperparameter selection. The target pretrained GNNs and test instances remain unchanged. Each entry reports Fidelity AUC / Robust Fidelity AUC, averaged over three seeds.

First, all BBBP configurations across GCN, GAT, and GIN are excluded from generator fitting and hyperparameter selection.

Target architecture on BBBPReFineSTAR, strict dataset holdout
GCN0.7796 / 0.77630.8494 / 0.8403
GAT0.7853 / 0.78450.8716 / 0.8671
GIN0.7526 / 0.75810.8185 / 0.8295

STAR exceeds target-fitted ReFine by 0.0698, 0.0863, and 0.0659 in Fidelity AUC on GCN, GAT, and GIN, respectively. The corresponding improvements in Robust Fidelity AUC show that the advantage also persists under input perturbations. These results demonstrate effective explanation transfer when the target dataset contributes neither to generator fitting nor to inference-hyperparameter selection.

Second, all GCN configurations across the six datasets are excluded from generator fitting and hyperparameter selection. The source pretrained GNNs comprise only GAT and GIN models.

Target dataset with GCNReFineSTAR, strict architecture holdout
MUTAG0.6181 / 0.61280.6393 / 0.6324
BBBP0.7796 / 0.77630.8497 / 0.8404
BACE0.5649 / 0.56280.5743 / 0.5722
HIV0.7166 / 0.72050.7136 / 0.7186
Benzene0.6674 / 0.64590.7823 / 0.7580
PROTEINS0.6224 / 0.62300.6564 / 0.6554

STAR achieves higher values on both metrics in five of the six datasets, with particularly substantial Fidelity AUC improvements on BBBP and Benzene, at 0.0701 and 0.1149. On HIV, STAR remains close to target-fitted ReFine, with differences of 0.0030 and 0.0019 on the two metrics. These results demonstrate transfer to an architecture family entirely absent from both generator fitting and hyperparameter selection.

Response to Weakness 2: Thank you for your question. The source-acquired explanatory prior provides relevance estimates for candidate construction and guides search proposals. Target-specific optimization then assesses these candidates using the target pretrained GNN’s predictive responses. Direct-transfer comparisons and generator-free controls examine the contribution of this guidance from complementary perspectives.

For direct transfer, the following comparison uses GIN and reports Fidelity AUC on the same BBBP target instances over three seeds.

Source dataset for ReFineTarget datasetSource-fitted ReFineTarget-fitted ReFineSTAR, dataset-held-out
BACEBBBP0.69010.76830.8176

Applying the ReFine explanation function fitted on BACE to the BBBP target GNN produces a 0.0782 decrease relative to fitting ReFine on the target setting. STAR achieves 0.8176 using a generator fitted across source settings excluding BBBP. This comparison illustrates the transfer mismatch and shows that STAR can exploit source-acquired explanatory information while adapting explanations to the target pretrained GNN without target-specific generator refitting.

For Cora, the requested Strict Local-only comparison gives the following results.

ArchitectureStrict Local-only Fidelity AUCSTAR Fidelity AUC
GCN0.896750.89697
GAT0.821490.82432
GIN0.889200.88920

The small differences are consistent with local optimization already recovering evidence with strong predictive support. The explanatory prior contributes when its relevance estimates and proposals identify useful candidates beyond those recovered locally. Its incremental benefit therefore depends on the target setting. On Cora, STAR preserves the strong performance of local optimization and provides a further improvement on GAT.

AuthorsOpenReview ↗

Rebuttal Part II

Response to Weakness 3: Thank you for your question. The reported 813 versus 348 forward calls count the complete explanation procedure, including local optimization, candidate construction, discrete search, and candidate assessment. The 2.34-fold ratio therefore does not correspond to 2.34 times as many discrete-search steps. STAR additionally constructs and evaluates explanatory-prior-guided candidates, which increases its total number of target-GNN calls.

This distinction explains the difference between the search settings and the overall inference cost. The additional evaluations assess whether source-guided candidates provide useful evidence for the queried target prediction. When local optimization already identifies strong evidence, the resulting fidelity differences can be small, as observed on Cora.

STAR uses a frozen explanatory prior generator across target settings, avoiding a separate generator-fitting procedure for each target GNN. The additional target-GNN evaluations assess source-guided and locally optimized candidates for the queried prediction, enabling explanation adaptation without refitting the explanatory prior generator.

Response to Weakness 4: Thank you for highlighting this case. The route-level Fidelity AUC results identify where the decrease occurs on Benzene–GIN.

SettingGeneratorLocal searchPrior-guided searchFinal STAR
Joint-fit0.72130.58090.68920.6223
Held-out0.70540.58380.58810.5998

The generator already provides strong predictive support in this configuration. Both searched routes have lower average Fidelity AUC, and the final assessment frequently selects these searched alternatives. The decrease therefore occurs despite the availability of useful source-acquired explanatory information.

The original procedure incorporates fusion during candidate construction by combining generator and locally optimized edge scores at different weights. These candidates undergo search and assessment under the target pretrained GNN. The final explanation is determined by this assessment.

Two features account for the observed behavior. First, the assessment criterion considers both prediction preservation and responses to evidence removal, whereas Fidelity AUC measures retained predictive support. These criteria can prefer different candidates. Second, the route comparison uses the local route as its reference, so a candidate is assessed relative to that route rather than required to improve on the generator’s Fidelity AUC.

The analysis identifies a specific failure mode in the original procedure. When the generator already provides strong retained predictive support, search and composite candidate assessment can replace it with an explanation that has lower Fidelity AUC. This accounts for the reported degradation and clarifies why target-adaptive refinement has different effects across configurations.

Response to Weakness 5: Thank you for your question. The evaluation examines predictive support and evidence consistency separately. Fidelity measures whether the retained evidence supports the target pretrained GNN’s prediction. Robust Fidelity measures predictive support after re-explanation of perturbed inputs. Repeated-run and perturbation-consistency analyses examine agreement between the selected explanations.

This distinction is important because different subgraphs can support the same prediction. Predictive support evaluates the relationship between the selected evidence and the target GNN’s response, while consistency evaluates whether similar evidence is recovered across runs or input changes. The supplementary analyses therefore provide information beyond confidence-based fidelity alone.

Evaluation across three seeds reduces dependence on a particular pretrained GNN checkpoint and captures variability across training runs. All explanatory methods use the same target checkpoints and test instances within each comparison. For single-class checkpoints, the fidelity values describe support for those predictors’ responses. Predictor behavior and explanation consistency provide the context for interpreting these values.

The practical usefulness of explanatory evidence also depends on users’ tasks and application contexts. Evaluation with real users in real-world settings is a central focus of our future research, together with further investigation of explanation stability.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.