{"id":"d607ab5f-3f79-45a4-bfed-53483fb50de0","arxiv_id":"2508.01925","paper_version":1,"verdict":"UNVERDICTED","confidence":"LOW","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":4,"one_line_summary":"GNN explanations become more robust if the model is retrained on edge-weighted graphs that match the explanation-time distribution, using an iterative shrinking-subgraph procedure.","lead":"This paper proposes an iterative framework that alternates between extracting explanation subgraphs from a graph neural network and retraining the network on weighted graphs built from those subgraphs. The claimed payoff is more robust explanations that stay reliable even when small, sparse subgraphs are requested.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Iterative retraining may change the model's predictions, so the claimed explanation-quality gains could be self-confirming rather than faithful to the original model.","rationale":"The reader's weakest assumption correctly identifies that retraining could distort the model's decision process. My concern sharpens this: the concrete failure mode is that the retrained model's predictions diverge from the original model, making the explanation quality metric a measure of self-consistency rather than faithfulness. This is load-bearing because the paper's core novelty is the iterative retraining loop; if the loop silently swaps the model being explained, the claimed improvement is not an improvement in explaining the original model. The abstract alone cannot rule this out, and the reader's UNVERDICTED verdict remains appropriate. The proposed test is direct: it checks prediction agreement and explanation faithfulness against the original model. If the paper already includes such an evaluation, the concern would be resolved; if not, the central claim is unsupported. No ad hominem is intended; the concern is about the experimental design and the interpretation of the metric.","tokens_in":743,"tokens_out":2588,"duration_ms":30239,"concrete_test":"On a benchmark with ground-truth explanations (e.g., synthetic motifs or known molecular functional groups), fix a GNN trained on unweighted graphs. Run the proposed iterative retraining. Then compute the retrained model's predictions on the original unweighted test set and measure agreement with the original model's predictions. Next, evaluate explanation quality with respect to the original model: fidelity (change in the original model's prediction when the explained subgraph is removed) and accuracy of the identified subgraph against ground truth. If agreement is below a chosen threshold (e.g., 95%) or fidelity/accuracy w.r.t. the original model is not better than non-retrained baselines, the improvement is self-confirming.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that retraining a GNN on importance-weighted graphs reduces distributional shift and improves explanation quality. For this to be meaningful, the retrained model must remain predictive of the original model's decisions on the original unweighted graphs; otherwise the final explanations describe a different model, and any measured improvement is an artifact of self-consistency. The abstract gives no evidence that predictions are preserved across iterations. In particular, training on progressively sparser weighted subgraphs could make the model rely on edge weights in ways that diverge from the original model, especially on non-explanatory edges. The paper's 'robustness' claim is only non-circular if explanation faithfulness is measured against the original model's outputs or against ground-truth motifs; if it is measured against the retrained model's own outputs, the iterative loop guarantees higher consistency but not higher fidelity. This is the load-bearing assumption: the iterative adaptation improves alignment without distorting the decision process.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an iterative framework for GNN explanation that alternates between identifying an explanation subgraph via soft mask optimization and retraining the GNN on edge-weighted graphs derived from that subgraph. Starting from a relatively large explanation subgraph and progressively shrinking it, the procedure aims to align the training data distribution with the weighted-graph distribution encountered during explanation, thereby improving explanation quality across different GNN backbones and explanation methods. The abstract claims consistent improvements on multiple benchmark datasets, but provides no equations, quantitative results, or details on the training schedule.","tokens_in":887,"tokens_out":2608,"duration_ms":29082,"significance":"If the iterative alignment genuinely improves the faithfulness of explanations to the original model without distorting its decision process, the framework would be a practical plug-in for existing GNN explainers. The idea of reducing distributional shift between training and explanation-time graphs is well-motivated and potentially applicable to any soft-mask-based explanation method. The promise of consistency across backbones and datasets is a useful empirical claim, and the framework's simplicity is a strength. However, the abstract alone does not establish that the retrained model remains faithful to the original model, which is the central risk of the approach.","major_comments":[{"comment":"The iterative loop in which edge weights are derived from explanation subgraphs and then used to retrain the GNN creates a feedback mechanism. If the evaluation of explanation quality is performed against the retrained model's own predictions, any improvement can reflect self-consistency rather than fidelity to the original model. The abstract does not state whether prediction accuracy on the original unweighted graphs is preserved across iterations, nor whether faithfulness is measured against external ground truth or the original model's outputs. This is load-bearing for the central claim of 'robust explanation quality,' and the manuscript should either provide such evidence or explicitly frame the method as producing an adapted model with a different decision process.","section":"Abstract"},{"comment":"The abstract reports consistent improvements across backbones, but it does not describe any control experiment that retrains the GNN on random or constant edge weights. Without such a control, it is impossible to attribute the improvement to the importance-aware weighting rather than to the increased capacity of the retrained model to fit sparse weighted graphs. The shrinking schedule and initial subgraph size are free parameters, and the abstract gives no ablation showing the method's sensitivity to them.","section":"Abstract"},{"comment":"The method's starting point is an explanation subgraph obtained by soft mask optimization on a relatively large subgraph, which the abstract asserts is reliable. If the initial mask is biased or misaligned with the original model's decision process, the iterative refinement will amplify that bias. The abstract offers no diagnostic or failure analysis for this scenario, so the robustness claim is conditional on an unverified premise.","section":"Abstract"}],"minor_comments":[{"comment":"The abstract uses the phrase 'the weighted graph distribution appeared during explanation'; the intended wording is likely 'the weighted graph distribution that appears during explanation.'","section":"Abstract"},{"comment":"The term 'explanation robustness' is not defined; it would be helpful to state whether it refers to stability across perturbations, fidelity to the original model, or consistency of extracted motifs.","section":"Abstract"},{"comment":"No error bars or statistical significance tests are mentioned; for a paper claiming consistent improvement across datasets, the authors should report variance across random seeds.","section":"Abstract"}],"recommendation":"major_revision","confidential_remarks":"The reviewer had access only to the abstract, so the empirical claims could not be independently verified. If the experiments evaluate fidelity against the retrained model rather than the original model, the paper would be more suitable for a workshop than a journal. Please ensure the full submission includes a control experiment with random edge weights and a clear statement of how faithfulness is measured relative to the original model."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is better than the abstract lets on. Most explanation methods optimize soft masks on a frozen model, and everyone knows the weighted graphs used at explanation time differ from the unweighted training graphs. Retraining the GNN on importance-weighted versions of its own explanation subgraphs, starting large and shrinking, is a sensible way to close that gap. It is a wrapper, not a new theory, but it could plausibly improve several existing explainers, and the claimed across-backbone consistency is worth taking seriously.\n\nThe paper also does something right in framing: it names the distributional shift explicitly and attacks it directly. The iterative scheme is concrete enough to implement from the abstract, which is more than many papers offer. If the full text shows the expected benchmarks and baselines, this could be a solid systems contribution.\n\nNow the soft spot, and it is the one that matters. Retraining on weighted graphs can change the model. If the final explanations are evaluated against the retrained model's own outputs, the loop guarantees consistency, not fidelity. The abstract gives no evidence that predictions are preserved across iterations, or that faithfulness is measured against the original model or ground-truth motifs. The stress-test note is right: this is the load-bearing assumption. The paper needs to show the adapted model still approximates the original on unweighted graphs, or evaluate against external ground truth. Without that, the robustness claim is self-confirming.\n\nOther issues are more minor. There are no equations, hyperparameters, or error bars visible, so I cannot audit the method or its results. But this is an abstract-only review; those may all be in the full text. The shrinking schedule and edge-weight rule are free parameters that could hide a lot of tuning, so the paper should report sensitivity.\n\nMy take: this deserves a serious referee. The idea is testable, the problem is real, and the potential payoff is broad. But the referee should push hard on the circularity question. I would not cite it yet, and I would only bring it to reading group if the full text shows an honest evaluation against the original model's decisions.\n\nRecommendation: send to peer review, with a referee who understands the difference between self-consistency and faithfulness.","headline":"Iterative explanation-model co-adaptation is a genuinely useful idea, but the abstract leaves the circularity worry unresolved.","tokens_in":1428,"tokens_out":1173,"would_cite":false,"duration_ms":15456,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"Training a GNN on the weighted graphs it is later asked to explain improves explanation quality.","keywords":["graph neural networks","explainability","distribution shift","edge masks","soft mask optimization","model retraining","iterative refinement","subgraph explanation"],"falsifier":"On a synthetic graph dataset with planted ground-truth edges, compare the final adapted model's explanation against the original model's one-shot soft-mask explanation: if the adapted model does not recover the planted edges at least as well, or if retraining degrades accuracy on the original unweighted graphs, the claim that distribution alignment is responsible for the gains would be refuted.","tokens_in":535,"feed_emoji":"🔁","tokens_out":4692,"duration_ms":53456,"temperature":0.7,"pith_summary":"This paper identifies a specific reason graph explanations fail: GNNs are trained on unweighted graphs, but most explanation methods ask them to score edges on graphs whose edge weights are continuous, so the model is effectively being interrogated on a distribution it never trained on. The paper claims that this distributional mismatch, not just weak mask optimization, is why sparse explanations are unreliable, and that closing the gap by retraining the GNN on weighted graphs built from its own explanations, starting large and shrinking, improves explanation quality across backbones. It offers a general iterative loop rather than a single new explainer.","feed_headline":"Retrain graph AI on weighted graphs for clearer explanations","feed_subtitle":"An explain–retrain–shrink loop closes the gap between binary training graphs and weighted explanation graphs.","key_machinery":"The iterative explanation–adaptation loop. The central object is the weighted graph formed by assigning importance-aware edge weights to explanatory and non-explanatory edges of the current subgraph. Retraining on that weighted graph carries the argument: it cancels the binary-to-continuous distribution shift that otherwise makes soft-mask gradients unreliable, allowing the subgraph to shrink in later rounds without losing fidelity.","core_discovery":"The central claim is that explanation quality improves when the model adapts to the weighted graphs used at explanation time. Concretely, the paper proposes alternating between identifying a soft-mask subgraph and retraining the GNN on a reweighted version of that graph, with explanatory edges upweighted and non-explanatory edges downweighted. Because soft masks are more reliable on larger subgraphs, the loop starts there and repeatedly shrinks the subgraph; each round retrains on the distribution it will next be asked to explain. The result is that the soft weighted graph becomes part of the model's training distribution, so gradients and importance scores are no longer computed out-of-distribution.","pith_inferences":["A natural test the paper leaves implicit is whether the retrained model's predictions on the original unweighted test graphs stay accurate; a model retrained on weighted graphs could drift from the deployment distribution.","The loop could plausibly transfer to node-level and edge-level explanation tasks, where the same binary-input/continuous-weight mismatch appears, but the paper's reported evidence is on benchmark graph tasks.","If the mechanism is truly distributional, the benefit should grow as the target subgraph shrinks; a useful experiment would plot explanation fidelity against subgraph size for the iterated method versus one-shot soft masks.","One subtle risk worth checking: retraining on weighted explanations can make the final model's decision process differ from the original model, so faithfulness must be measured against whichever model is ultimately deployed."],"forward_implications":["Soft-mask explanation methods can be improved without changing their loss or architecture, just by retraining the model on the weighted graphs they produce.","Sparse, small-subgraph explanations become reliable, which matters because users typically want the smallest sufficient evidence for a prediction.","The same loop can be dropped into different GNN backbones and different explanation methods, making it a plug-in rather than a replacement.","The paper reports consistent gains across benchmark datasets and backbones, so the loop generalizes across architectures."],"supporting_citations":[],"fun_headline_variants":["Retrain GNNs on weighted graphs to fix explanation drift","Iterative reweighting loop stabilizes graph explanations","Explain, retrain, shrink: robust graph explanations","Adapt GNNs to explanation graphs for better masks","From binary to continuous: robust graph explainer"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole loop works only if the soft edge weights from the initial large-subgraph explanation are trustworthy enough that retraining on them sharpens the model rather than confirming a wrong signal.","fun_headline_variants_meta":{"raw":{"variants":["Retrain GNNs on weighted graphs to fix explanation drift","Iterative reweighting loop stabilizes graph explanations","Explain, retrain, shrink: robust graph explanations","Adapt GNNs to explanation graphs for better masks","From binary to continuous: robust graph explainer"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000189,"raw_usage":{"total_tokens":1314,"prompt_tokens":904,"completion_tokens":410,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":520,"completion_tokens_details":{"reasoning_tokens":333}},"tokens_in":520,"tokens_out":410,"duration_ms":4777,"temperature":1.0,"reasoning_tokens":333,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T05:16:58.689288+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a synthetic graph dataset with planted ground-truth edges, compare the final adapted model's explanation against the original model's one-shot soft-mask explanation: if the adapted model does not recover the planted edges at least as well, or if retraining degrades accuracy on the original unweighted graphs, the claim that distribution alignment is responsible for the gains would be refuted.","supporting_citations":[],"review_version":1}