{"id":"e2bb93fe-7232-43ed-aa26-997758e3df3d","arxiv_id":"2506.12378","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Component-level SHAP and ALE analysis via state-fidelity pseudo-models reveals differing feature importance across feature maps, ansatze, quantum kernels, and decision functions in a QML classifier.","lead":"This preprint applies classical explainability tools, SHAP and ALE, to individual components of hybrid quantum machine learning models by measuring quantum state fidelity instead of model predictions. It shows that different components emphasize different input features, which could help engineers debug and tune quantum classifiers.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 'Yes' conclusion rests on an unvalidated pseudo-model: fidelity to a class-average quantum state is used as the model output, but no evidence shows this reference similarity tracks the VQC/QSVC decision function, so the SHAP/ALE plots may explain the reference, not the model.","rationale":"The reader's conditional verdict already identifies the proxy problem, and my independent reading points to the same load-bearing assumption, sharpened into a concrete definitional ambiguity. The paper's headline contribution is a framework; the framework's two adaptations both route through a pseudo-model that measures fidelity to class-average states. Since the actual QML models use different decision mechanisms, the mapping from fidelity to prediction is nontrivial and is nowhere validated. This is distinct from a simple reproducibility complaint: even with perfect code, the plots would still be about similarity to a constructed reference unless the proxy is tested. I therefore agree with the reader's weakest_assumption. A single validation step, rank-ordering pseudo scores against true decision scores, would settle whether the concern lands. If it does land, the paper should be revised to either defend and validate the proxy or soften the central claim to 'component sensitivity analysis' rather than 'explainability of QML models.' Because the reader already conditions acceptance on this validation, the verdict need not change; the concrete test is the missing gate.","tokens_in":10734,"tokens_out":3849,"duration_ms":52204,"concrete_test":"On the held-out test set used in Section V, compute the pseudo-model fidelity scores that feed the SHAP/ALE analyses (e.g., fidelity of each encoded state to the class-average reference, and fidelity of perturbed ansatz outputs to the trained ansatz state) and the actual model decision scores (QSVC decision function values or VQC output probabilities). Compute the Spearman rank correlation between the two score vectors and the AUC of the pseudo scores against the true labels; also compare with the correlation obtained when the class-average reference is replaced by a random fixed quantum state. If the pseudo-model AUC is near chance or the rank correlation with actual decisions is weak, the component explanations do not explain the model's predictions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV introduces pseudo-models that replace the model prediction with state fidelity to a 'class average quantum state,' and Section V interprets the resulting SHAP and ALE plots as component-level explanations. The central claim ('The answer: Yes') requires that this fidelity score is a faithful surrogate for what the model actually computes. That condition is the least secure part of the argument. The paper never defines the class-average state. If it is a normalized sum of statevectors, cancellation can make it unrepresentative or even near-orthogonal to most class members; if it is a density-matrix mixture, the pseudo-model output becomes an average squared overlap with the class, which is not the QSVC decision function (a kernel-weighted sum over support vectors) and not the VQC output (a parameterized measurement expectation). A fidelity score can be perfectly well-defined and still have almost no relationship to the model's prediction. The plots in Figures 6-12 therefore show sensitivity of a reference-similarity measure, and the paper provides no correlation, calibration, ablation, or comparison to the actual decision function to show that this measure explains the model. This is a correctness risk, not just a missing artifact: the compositional inference from 'component explains reference similarity' to 'component explains model' is unsupported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a modular framework for explainable quantum machine learning: instead of explaining a trained VQC or QSVC as a whole, it applies SHAP and ALE to individual components (feature map, ansatz, quantum kernel, SVC decision function) via 'pseudo-models' whose output is state fidelity to class-average quantum states. The framework is demonstrated on the Pima Indians Diabetes dataset, with waterfall plots, ALE curves, kernel eigenvalue/PCA analysis, and a surrogate-based SHAP analysis of the SVC decision function. The authors conclude that component-based explainability reveals feature and parameter importance that monolithic explainability would hide, answering their stated research question in the affirmative.","tokens_in":10926,"tokens_out":4504,"duration_ms":57821,"significance":"If the central assumption is validated, the modular decomposition is a plausible and potentially useful direction for XQAI: component-level attribution could guide feature-map and ansatz tuning in ways that monolithic explanations cannot. The paper has concrete strengths: it uses a real-world dataset, adapts two model-agnostic explainability methods, performs a surrogate analysis of the QSVC decision function with a reported R² of 0.98, and includes spectral analysis of the quantum kernel. However, the load-bearing claim depends on an unvalidated pseudo-model, and the evaluation is almost entirely qualitative, with no model accuracy numbers, no comparison to classical baselines despite a promise to do so, and no error bars or repeated-run statistics. The contribution is therefore promising but currently not established at the level required for a journal publication.","major_comments":[{"comment":"The pseudo-model that replaces the model prediction with fidelity to a 'class average quantum state' is never formally defined. If the class-average state is a normalized sum of statevectors, destructive interference can make it unrepresentative or nearly orthogonal to most class members; if it is a density-matrix mixture, the pseudo-model output becomes an average squared overlap, which is not the QSVC decision function (a kernel-weighted sum over support vectors) nor the VQC output (a parameterized measurement expectation). All SHAP and ALE plots in Figures 6-12 therefore explain similarity to a constructed reference, not necessarily the model's decision. The authors must provide evidence that the fidelity score tracks the actual decision function, for example by reporting the correlation or calibration between pseudo-model outputs and VQC/QSVC outputs on validation samples, or by an ablation showing that known decision-relevant features are recovered.","section":"Section IV and Section V.A"},{"comment":"The paper promises 'a comprehensive and in depth analysis into the values will occur against classical ML algorithms', but no classical ML baseline results are reported, and no accuracy or F1 scores are given for the VQC or QSVC models themselves. Without these numbers, the conclusion that component-based explainability shows 'benefits' and 'the need' for the approach is not quantitatively supported. Please add a results table with VQC/QSVC and classical model metrics across repeated runs or seeds, and include quantitative explanation-quality measures such as agreement with the actual decision function or stability of SHAP values across runs.","section":"Section III.E and Section V"},{"comment":"The ansatz SHAP adaptation is described only as 'slightly perturbing the original ansatz weights', but the perturbation scale, the number of SHAP samples, the background dataset, and the exact baseline for comparison are not specified. Likewise, the 'central sample from each class' used for kernel SHAP is undefined, and the ALE binning parameters are not given. These details are needed for reproducibility and for interpreting the parameter-importance plots in Figures 11 and 12. Please specify all hyperparameters of the explainability adaptations, or release the code in a way that makes the exact procedures available.","section":"Section IV.B and Section V.B.1"}],"minor_comments":[{"comment":"Figure references are inconsistent: the ZZ feature map is called 'Figure 1' in the text but the actual circuit diagram is Figure 2, and several references appear as 'Figure ??' (e.g., for the ansatz circuit and the bloch sphere). Please correct these cross-references.","section":"Section II.C.1 and Section V.A"},{"comment":"There are numerous typos and grammatical slips that should be fixed, including 'alforithms' (Section II), 'od' (Section II.C), 'helper' (Section V.C.1), 'paperion' (Section V.C.1), and 'affect' where 'effect' is meant in the ALE description (Section II.H.2).","section":"Throughout"},{"comment":"The sentence 'lit review table' appears to reference a table that is not present in the manuscript; either add the table or remove the dangling reference.","section":"Section III.E"},{"comment":"References [2] and [7] are duplicate entries for the same Gil-Fuster et al. paper, and some citations (e.g., [12] for the VQC diagram) do not appear to match the cited content. The bibliography should be cleaned and checked against the text.","section":"References"},{"comment":"The claim that these are 'the first component based explainability techniques' is not supported by the related-work survey, which does not systematically demonstrate absence of prior component-level work. Please temper the claim or provide a more thorough literature search.","section":"Section VI"}],"recommendation":"major_revision","confidential_remarks":"The manuscript reads more like an extended project report than a finished journal article, but the core idea is viable and could become publishable after substantial revision. The main gate is validation of the fidelity-based pseudo-model: without a quantitative link to the actual VQC/QSVC decision functions, the central 'Yes' conclusion is not established. I would also ask the authors to verify the novelty claim against a broader literature search and to include the promised classical ML comparisons."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: the paper's decomposition of QML models into feature map, ansatz, kernel, and decision function is a genuinely new angle, and the adapted SHAP/ALE plots are internally consistent. But the central 'Yes' conclusion rests on an unvalidated pseudo-model: fidelity to a class-average quantum state is used as the model output, and there is no evidence this reference similarity tracks the VQC/QSVC decision function. Without that, the SHAP/ALE plots may explain the reference, not the model.\n\nWhat's new: prior work applies SHAP/ALE to whole QML models or Shapley values to gates; the component-wise decomposition with fidelity-based pseudo-models is a legitimate step. The paper also includes kernel spectral analysis and a surrogate for the SVC decision function, which are sensible complementary views. The related work is honestly surveyed and clearly differentiated.\n\nWhere it's soft: the stress-test note is right. The class-average quantum state is never defined; a normalized sum of statevectors and a density-matrix mixture behave very differently. The paper provides no correlation, calibration, ablation, or comparison to the actual decision function to show the pseudo-model is a faithful surrogate. The comparison against classical ML baselines promised in Section III.E is missing. There are no accuracy or F1 numbers, no error bars across runs, and figures are often referenced with placeholders. The writing has typos and the 'lit review table' is an obvious leftover.\n\nProportionately: the idea is plausible and worth exploring, but the current evidence is illustrative, not confirmatory. The compositional inference from 'component explains reference similarity' to 'component explains model' is unsupported. That is the load-bearing flaw, and it is fixable: define the class-average state, validate the fidelity score against the actual decision function on held-out data, and include baselines and error bars.\n\nThe paper is for QML practitioners interested in debugging and feature engineering. A serious referee should engage with it if the authors are willing to revise heavily; the novelty supports referee time, but the current form would need major changes before acceptance.","headline":"A plausible component-level decomposition for QML explainability, but the central claim rests on an unvalidated fidelity proxy and needs major revisions.","tokens_in":11545,"tokens_out":2546,"would_cite":false,"duration_ms":25347,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":["03.67.Lx"],"model":"deepseek-v4-flash","headline":"This paper answers yes to whether QML can be made explainable by explaining its core components, using fidelity-based SHAP and ALE to reveal feature and parameter importance that monolithic analysis hides.","keywords":["quantum machine learning","explainability","SHAP","ALE","component-based framework","quantum state fidelity","variational quantum classifier","quantum support vector classifier"],"falsifier":"Take a trained classifier and add a synthetic feature that is random noise but happens to push encoded states toward the class-average reference; if fidelity-based SHAP reports that noise feature as important while the model's accuracy is unchanged when the feature is dropped, the fidelity-to-class-average assumption fails. In other words, compare component-level SHAP importance with leave-one-feature-out model accuracy: any feature that is important in the fidelity plots but has no effect on accuracy would falsify the link between fidelity and prediction.","tokens_in":10433,"feed_emoji":"🧩","tokens_out":6840,"duration_ms":73159,"temperature":0.7,"pith_summary":"This paper sets out to answer a yes-or-no question: can a quantum machine learning model be made explainable by explaining its constituent components instead of the whole model? It answers yes, by replacing the model's prediction with a quantum state fidelity measure and feeding that into two classical explainability tools, SHAP and ALE, separately for the feature map, the variational ansatz, the quantum kernel, and the classical decision function. Applied to a variational quantum classifier and a quantum support vector classifier on the Pima Indians diabetes dataset, the method surfaces component-specific feature and parameter importance, such as Age and BMI mattering mainly at the decision stage while Pregnancies matters mainly at the initial encoding stage. The paper positions this as the first component-based explainability framework for quantum machine learning.","feed_headline":"Quantum ML becomes explainable when taken apart","feed_subtitle":"Fidelity-based SHAP and ALE reveal which features matter at each step, from encoding to decision.","key_machinery":"The central mechanism is the 'pseudo model': a wrapper that takes a quantum state (or a state produced by perturbed ansatz parameters) and returns its fidelity $F = |\\langle \\psi | \\phi \\rangle|^2$ to a reference state, typically the average encoded state of a class. This turns quantum states into scalar outputs that classical model-agnostic explainers such as SHAP (Shapley additive explanations, a game-theoretic feature attribution method) and ALE (accumulated local effects, a perturbation-based feature effect plot) can consume, at which point those methods compute Shapley values or accumulated local effects over the original input features for the feature map and kernel, or over the ansatz parameters for the variational circuit. For the quantum kernel, the technique instead evaluates kernel values against a central sample per class, and for the decision function a surrogate model is trained on the SVC decision output with an $R^2$ of 0.98.","core_discovery":"On the paper's own terms, the central discovery is that explainability can be inferred for a QML algorithm by decomposing it into components and applying explainability techniques to each component independently. The key adaptation is the use of 'pseudo models' that output quantum state fidelity to a class-average state, so that SHAP and ALE, which normally require a scalar prediction, can be applied to quantum states and to ansatz parameters. The empirical results show that feature importance differs across components in a way that is invisible to monolithic analysis: for example, Glucose and SkinThickness retain similar importance across stages, but Age and BMI gain importance only in the decision function, while Pregnancies matters in the initial encoding but not in the final decision. The paper concludes that the answer to the question of whether component-based explainability works is yes, and claims this is the first such component-based technique.","pith_inferences":["Because the fidelity reference is built from class averages, the technique may conflate 'close to a typical example' with 'increases the predicted probability'; replacing the reference with a probability-weighted mixture of class states could separate these two notions.","A natural extension is to apply the same component decomposition to quantum regression, using target-value-conditioned reference states and inspecting how feature effects on the output emerge across components.","The component-level plots could be used to compare two feature maps directly: if one feature map assigns high importance to the same features in the decision function as the classical model, it may be a better inductive bias for that task.","The fidelity-based SHAP values for a single sample are local explanations; aggregating them across many samples could produce a global component-importance map that would let practitioners see, before training, which features the encoding will amplify."],"forward_implications":["Feature engineering can now be targeted: features that matter only at the decision stage might be kept even if they look weak in the encoding, and features that matter only at encoding might be worth transforming or scaling.","Model tuning can become layer-specific: parameter-level SHAP values identify which ansatz layers or repetitions drive the output, enabling pruning or reinitialization of the least important parameters.","The modular framework is intended to be model-agnostic, so the same pseudo-model approach can be dropped into other QML algorithms and other components, such as different optimizers or quantum-classical loop stages.","Component-level importance can serve as a diagnostic for where a model is failing: if the encoding stage already separates the classes, the decision stage is the bottleneck, and vice versa."],"supporting_citations":[{"why":"Defines SHAP values as the game-theoretic feature attribution used for the feature map, ansatz, and decision-function explanations.","marker":"[20]"},{"why":"Defines accumulated local effects (ALE), the perturbation-based method adapted to the feature map via fidelity pseudo-models.","marker":"[22]"},{"why":"Argues that classical explainability techniques can be incompatible with parameterised quantum circuits, motivating the component-wise approach.","marker":"[7]"},{"why":"Earlier monolithic feature-importance study on quantum models that the paper extends by decomposing the algorithm.","marker":"[24]"},{"why":"Earlier SHAP-on-quantum-classifier study that lacked class-comparison and classical baselines; the paper expands on it.","marker":"[25]"},{"why":"Explains quantum circuits with Shapley values at the gate level but does not cover feature maps, kernels, or optimizers.","marker":"[27]"},{"why":"Supplies the quantum state fidelity formula used as the output of the pseudo models.","marker":"[29]"}],"fun_headline_variants":["Splitting quantum ML reveals feature roles per stage","Component-wise SHAP and ALE explain quantum models","Fidelity-based pseudo models make QML explainable","Decomposing QML gives stage-specific feature insights"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The entire pipeline assumes that the fidelity of a quantum state to a class-average reference state accurately represents how that component contributes to the model's prediction; if fidelity measures similarity to a constructed average rather than decision-relevant signal, the resulting plots are not explanations of the model.","fun_headline_variants_meta":{"raw":{"variants":["Splitting quantum ML reveals feature roles per stage","Component-wise SHAP and ALE explain quantum models","Fidelity-based pseudo models make QML explainable","Decomposing QML gives stage-specific feature insights"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00023,"raw_usage":{"total_tokens":1464,"prompt_tokens":911,"completion_tokens":553,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":527,"completion_tokens_details":{"reasoning_tokens":490}},"tokens_in":527,"tokens_out":553,"duration_ms":6537,"temperature":1.0,"reasoning_tokens":490,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T00:52:28.024323+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a trained classifier and add a synthetic feature that is random noise but happens to push encoded states toward the class-average reference; if fidelity-based SHAP reports that noise feature as important while the model's accuracy is unchanged when the feature is dropped, the fidelity-to-class-average assumption fails. In other words, compare component-level SHAP importance with leave-one-feature-out model accuracy: any feature that is important in the fidelity plots but has no effect on accuracy would falsify the link between fidelity and prediction.","supporting_citations":[{"cited_title":"Lundberg and Su-In Lee","cited_arxiv_id":null,"evidence_quote":"Defines SHAP values as the game-theoretic feature attribution used for the feature map, ansatz, and decision-function explanations."},{"cited_title":"Study of Feature Importance for Quantum Machine Learning Models","cited_arxiv_id":"2202.11204","evidence_quote":"Earlier monolithic feature-importance study on quantum models that the paper extends by decomposing the algorithm."},{"cited_title":"eXplainable AI for Quantum Machine Learning","cited_arxiv_id":"2211.01441","evidence_quote":"Earlier SHAP-on-quantum-classifier study that lacked class-comparison and classical baselines; the paper expands on it."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Explains quantum circuits with Shapley values at the gate level but does not cover feature maps, kernels, or optimizers."},{"cited_title":"Qubit fidelity, 2025","cited_arxiv_id":null,"evidence_quote":"Supplies the quantum state fidelity formula used as the output of the pseudo models."}],"review_version":1}