{"id":"9f6e9f07-7fba-47e0-bd1f-29c126d727b1","arxiv_id":"2411.18141","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"On a 32-point water quality dataset from Durban, a quantum support vector classifier reached 75% accuracy while a quantum neural network consistently failed to train.","lead":"This paper applies two quantum machine learning models, a quantum support vector classifier and a quantum neural network, to water quality data from 32 sites in Durban, South Africa. It reports the classifier reaches 75 percent accuracy while the quantum neural network fails to learn.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The central QSVC-vs-QNN comparison is not supported because all reported metrics appear to be computed on the oversampled 32-point dataset with no held-out split; with only three positive samples, 0.75 accuracy likely reflects training-set memorization rather than generalization.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: the 32-point dataset, with only 3 positive samples before oversampling and no train/test split, cannot support the reported accuracy comparison. I agree with the reader's REJECT verdict. The paper has some honest self-assessment—it acknowledges the lack of novelty and the QNN's failure—and the application domain is unusual. However, the central empirical claim depends on metrics that are not demonstrably out-of-sample. If the metrics are in-sample on an oversampled set, they are inflated and non-comparable across models, and the QSVC/QNN comparison collapses. No classical baseline is needed to see this; the issue is internal validity. The proposed test—spatial cross-validation with oversampling confined to training folds—would directly settle whether QSVC genuinely generalizes better on this catchment than QNN. Until such an evaluation is provided, the abstract's comparative claim is unsupported, and the REJECT verdict stands.","tokens_in":7519,"tokens_out":2380,"duration_ms":23448,"concrete_test":"Re-run the experiment with leave-one-location-out or 5-fold spatial cross-validation on the original 32 unoversampled samples, applying random oversampling only inside each training fold and never to the test fold; report QSVC and QNN accuracy with 95% confidence intervals and per-fold results. If the QSVC accuracy drops to near chance or its confidence interval overlaps the QNN's, the central claim does not survive.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The strongest claim—that QSVC is easier to implement and yields higher accuracy than QNN (Abstract; Tables 1 and 2)—rests entirely on the quality of the reported accuracy numbers. The load-bearing flaw is the evaluation protocol. Section 4 states the data comprise 32 points, with only 3 'acceptable' versus 29 'unacceptable' water-quality measurements, and that the predictor was balanced using random oversampling. No train/test split, spatial cross-validation, or error bars are reported anywhere in the paper, and the Data Availability and Code Availability sections are both N/A. If the metrics in Tables 1 and 2 are computed on this oversampled in-sample set, they are not estimates of generalization: random oversampling simply duplicates the three positive instances, so a classifier can obtain high accuracy and perfect recall by memorizing majority patterns, while the QNN's near-chance losses and constant 0.4996-0.5000 loss values indicate it is not fitting even that same set. The paper itself concedes that the QNN is 'inherently flawed' and that future work with more data points is needed. A secondary inconsistency reinforces the concern: Section 4.2 attributes the QNN failure partly to 'the ReLU function in the intermediate layers,' but the QNN theory in Section 3.2 describes a variational quantum circuit with rotation gates and measurements, with no ReLU layers specified; this makes the QNN comparison ill-defined. Without a held-out evaluation on the original 32 samples, the central claim that QSVC outperforms QNN is unsubstantiated.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper applies two quantum machine learning methods, a quantum support vector classifier (QSVC) and a quantum neural network (QNN), to classify 32 water-quality measurement points in the Umgeni catchment (U20A) as acceptable or unacceptable for recreation based on an E. coli threshold. The authors report QSVC accuracies of 0.5833 (linear) and 0.7500 (polynomial and RBF kernels), and QNN accuracies between 0.4167 and 0.5000 with near-constant loss around 0.5, concluding that the QSVC is easier to implement and yields higher accuracy. The paper contains no train/test split, no error bars, no classical baseline, and no data or code availability.","tokens_in":7849,"tokens_out":3461,"duration_ms":31397,"significance":"If the results were properly validated, the paper would provide a useful application of QML to environmental water-quality monitoring, an area with limited prior QML work. The authors are candid about the QNN's failure and about the small dataset. However, the central comparison between QSVC and QNN rests on metrics that are not shown to estimate generalization, and the QNN model itself is not sufficiently specified. The paper ships no reproducible artifacts (Data availability: N/A, Code availability: N/A), which further weakens its contribution.","major_comments":[{"comment":"No train/test split, spatial cross-validation, or any held-out evaluation is reported; the metrics in Tables 1 and 2 appear to be computed on the same 32-point dataset that was balanced by random oversampling. With only 3 original positive samples, random oversampling duplicates the minority class, so the perfect recall (1.0000) and F1/precision values in Table 1 can reflect memorization of duplicated training points rather than generalization. This is load-bearing for the abstract's claim that the QSVC 'yields a higher accuracy.' Please report evaluation on a held-out test set or repeated stratified cross-validation, with per-fold metrics and error bars.","section":"Section 4, Tables 1 and 2"},{"comment":"The QNN failure analysis cites ReLU 'dead neuron' problems, Xavier weight initialization with fan-in/fan-out, and 'intermediate layers,' but Section 3.2 defines a QNN as a variational quantum circuit with rotation gates, a unitary U(θ), and a measurement (Eqs. 5-7), with no ReLU activations or classical intermediate layers. The actual trained model is therefore ill-defined, and the comparison in Table 2 is not interpretable. Specify the exact circuit ansatz, number of qubits, feature-encoding scheme, observable, and whether any classical layers with ReLU were included.","section":"Section 4.2 vs. Section 3.2"},{"comment":"The noise models (depolarizing noise, amplitude damping) are described mathematically, but no simulation details are given regarding where the noise is inserted, the number of shots, the simulator backend, or the qubit count, and only a single run is reported per configuration. The near-constant loss values (0.4996-0.5000) across epochs and optimizers are thus not diagnostic: they are consistent with a model that always predicts the majority class after oversampling, which would also explain the reported accuracy of 0.5000 and all-zero F1/precision/recall. Multiple random seeds and loss curves are needed to support any ranking among optimizers.","section":"Section 4.2, Eqs. (10)-(11) and Table 2"},{"comment":"The introduction asserts that QML 'can achieve comparable or superior accuracy to classical methods,' but the paper explicitly declines to build any classical baseline (Section 5), calling such a comparison 'unwarranted.' The reported 0.75 QSVC accuracy has no reference point, so the absolute performance and any claim of superiority over classical methods are unsupported. Either remove the introduction's claim or add a classical SVM trained on the same features and evaluation protocol.","section":"Section 1 and Section 5"}],"minor_comments":[{"comment":"The abstract contains a typo ('accraucy' should be 'accuracy'), and the phrase 'weight initializations were considered' is grammatically incomplete.","section":"Abstract"},{"comment":"The sentence 'Both kernels perform similarly well' is ambiguous because the preceding discussion contrasts the linear kernel with the polynomial and RBF kernels; please clarify that 'both' refers to polynomial and RBF.","section":"Section 4.1, text after Table 1"},{"comment":"The column header 'Architectural Components' actually lists hyperparameters (learning rate, optimizer, noise probabilities); rename the column to 'Model Configuration' or similar.","section":"Table 2"},{"comment":"The amplitude-encoding sum is written as (P_{2^n-1}_{i=1} x_i |i>) with the normalization condition; the summation index should start at 0, and the normalization should be enforced on the coefficients, not on the state itself as written.","section":"Eq. (5)"},{"comment":"The publisher locations for Schuld's books are listed as '???' and should be completed.","section":"References [25] and [26]"},{"comment":"The Data availability and Code availability both say 'N/A', making the reported results impossible to reproduce or verify; at minimum, list the exact features used, the preprocessing pipeline, the oversampling procedure, and the software library employed.","section":"Declarations"}],"recommendation":"reject","confidential_remarks":"The central comparison in the abstract is not supported by the reported evaluation protocol, and the QNN architecture is insufficiently specified to be reproduced. The load-bearing issues (no held-out evaluation, no reproducible code/data, internal inconsistency between the QNN theory and the QNN failure analysis) cannot be corrected by text revisions alone; they require new experiments and a properly designed evaluation. I also note that the paper contains several self-referential promotional passages and a self-citation in the future-work section; these do not affect the technical assessment but are inappropriate for a journal submission."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is an honest, clearly written application of standard QSVC and QNN methods to a small water-quality dataset, and it explicitly disclaims novelty. But the evaluation protocol is not adequate to support the headline claim that QSVC beats QNN. I agree with the reader's REJECT, and the stress-test note lands: the metrics in Tables 1 and 2 appear to be computed on the oversampled 32-point set with no held-out split, so they measure memorization rather than generalization.\n\nWhat the paper does well: it is candid about its limitations—data availability is N/A, code is N/A, the authors acknowledge the methods are not novel, and they note the QNN is 'inherently flawed' and that future work with more data points is needed. The writing is clear and the choices are explained. The application domain is genuinely new.\n\nThe soft spots are load-bearing. First, no train/test split, no cross-validation, no error bars. With 3 positive samples and random oversampling, 0.75 accuracy is exactly what you'd expect from majority-class memorization. Second, the QNN comparison is ill-defined: Section 4.2 blames the 'ReLU function in the intermediate layers' for the dead neuron problem, but Section 3.2 describes a variational circuit with rotation gates and measurements, with no ReLU layers anywhere. That is a real inconsistency. Third, the paper's refusal to compare to classical baselines, defended as comparing 'apples to bananas,' is a stance that reduces the practical value of the study, though it is not dishonest.\n\nThe paper does not deserve a serious referee in its current form. The missing evaluation protocol is a fundamental flaw that would have to be fixed by collecting more data and running proper held-out evaluation, which is beyond a simple revision. I would not cite it, and I would not bring it to reading group. If the authors redo the evaluation on a larger dataset with classical baselines, the application might be worth another look. Desk reject, with a note to the authors that the topic is interesting and the honesty is appreciated.","headline":"Honest small-case QML application, but the missing train/test split and an ill-defined QNN make the QSVC-vs-QNN comparison unsupported.","tokens_in":8394,"tokens_out":2446,"would_cite":false,"duration_ms":20039,"reading_group":"no","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper reports that a quantum support-vector classifier reaches 75% accuracy on a 32-point water-quality dataset from Durban, outperforming every quantum neural network variant it tested.","keywords":["Water Quality Prediction","Quantum Machine Learning","Quantum Support Vector Classifier","Quantum Neural Network","Dead Neuron Problem","Class Imbalance","Umgeni Catchment","Escherichia coli"],"falsifier":"Run leave-one-out cross-validation on the original 32 measurements, keeping the three acceptable samples intact, and report accuracy for each fold; if the QSVC's average held-out accuracy is at or near 50%, or if it never predicts the acceptable class, the claimed QSVC advantage over the QNN is not established.","tokens_in":7300,"feed_emoji":"💧","tokens_out":8402,"duration_ms":70757,"temperature":0.7,"pith_summary":"The paper applies two quantum machine learning methods to a small field dataset from the Umgeni catchment (U20A) in Durban, where 32 sampling locations were labeled acceptable or not acceptable for recreation by an E. coli threshold of 235 MPN/100 mL. Its central claim is that the quantum support-vector classifier (QSVC) is easier to implement and more accurate than the quantum neural network (QNN) on this data: the polynomial and RBF quantum kernels both reach 75% accuracy, while the best QNN variant reaches 50%. The QNN repeatedly collapsed into a dead-neuron state in which every output was constant, so only accuracy and loss could be compared. The authors position the work as a new application domain for QML rather than a new algorithm, and they deliberately do not compare against classical models.","feed_headline":"Quantum SVM beats quantum neural net on Durban water data","feed_subtitle":"On 32 Umgeni samples, the support-vector classifier hits 75% accuracy while every neural-net variant stalls near 50%.","key_machinery":"The load-bearing object is the quantum kernel $K(x_i,x_j)=|\\langle\\psi(x_i)|\\psi(x_j)\\rangle|^2$: classical water-quality features are encoded into quantum states by a variational feature map, the kernel measures state overlap, and classical SVM machinery then finds the separating hyperplane. For the QNN, the central object is the parametrized variational circuit $U(\\theta)$ with angle or amplitude encoding, trainable rotation gates, and measurement of an observable $\\langle\\hat{O}\\rangle$; its gradient-based training is what collapsed into the dead-neuron regime. The paper's comparison is therefore between a kernel method whose quantum part is a fixed feature map and a variational circuit whose parameters must be optimized.","core_discovery":"On the paper's own terms, the discovery is that a quantum kernel method can classify recreational water quality from E. coli measurements more reliably than a variational quantum neural network can. Using a quantum feature map to encode the field measurements into quantum states, the QSVC with polynomial and RBF kernels produced identical metrics, including accuracy 0.7500, F1 0.8000, and perfect recall; the linear kernel scored lower. The QNN, tested across optimizers, learning rates, weight initializations, and noise models, either stayed at loss near 0.4996 or drifted from 0.4167 to 0.5000 accuracy, and its precision, recall, and F1 were all 0 because of the dead-neuron failure. The paper concludes that for this dataset the QSVC is the pragmatic choice, and that the raw data alone is insufficient for an effective QNN without feature engineering.","pith_inferences":["Not stated in the paper but likely: because no held-out test set or cross-validation is reported, the 0.75 QSVC accuracy is an in-sample figure and the true generalization error on new Umgeni sites could be substantially lower.","The perfect recall and lower precision of the linear kernel suggest the model almost always predicts the majority class; after random oversampling of only three acceptable samples, minority-class performance estimates are fragile.","A natural extension the paper leaves implicit is a spatial split: train on some monitoring sites, test on others, to see whether the QSVC advantage survives geographic extrapolation.","Since the paper explicitly refuses classical baselines, a skeptical reader could test a classical SVM with the same kernels on the same 32 points; if it matches 0.75, the quantum encoding adds nothing measurable on this dataset."],"forward_implications":["If QSVC is the better tool, practitioners building small, imbalanced environmental datasets can start with a quantum kernel classifier rather than a variational QNN.","The identical polynomial and RBF performance suggests the chosen feature map makes the classes separable in a way that is insensitive to kernel shape; tuning degree or $\\gamma$ is the next test.","The dead-neuron failure means QNN results on raw tabular data should be reported with caution; without feature engineering, precision and recall can be undefined even when accuracy is near 50%.","The paper's 0.75 accuracy is the reported number to beat for future quantum models on the U20A recreation-use dataset."],"supporting_citations":[{"why":"Supplies the supervised-learning-with-quantum-computers background that defines the QML task and notation.","marker":"[25]"},{"why":"Supplies the quantum machine learning formalism, including feature maps and quantum kernels, used to implement QSVC and QNN.","marker":"[26]"},{"why":"Provides the closest analogous real-world water-domain QML application, flood forecasting on the Wupper River, that motivates the study design.","marker":"[30]"},{"why":"Supplies the quantum-train LSTM/QNN flood-prediction approach used as a contrast to the paper's simpler QNN setup.","marker":"[31]"},{"why":"Reports QSVC and QNN applications to crop yields from water and soil data, against which the paper positions its own domain novelty.","marker":"[32]"},{"why":"Reports a hybrid quantum deep-learning rice-yield forecast, another QML environmental application used to argue the application space is active.","marker":"[33]"}],"fun_headline_variants":["Quantum SVM beats quantum neural net on Umgeni water data","QSVC outperforms QNN for water quality in South African catchment","Quantum kernel classifier wins over variational net on E. coli data","75% accuracy: quantum SVM trumps quantum neural network","Support vector quantum classifier defeats neural net on water samples"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole comparison rests on the assumption that 32 spatially scattered field samples, with only 3 'acceptable' readings before random oversampling, represent the U20A catchment well enough for classifier accuracy to mean anything.","fun_headline_variants_meta":{"raw":{"variants":["Quantum SVM beats quantum neural net on Umgeni water data","QSVC outperforms QNN for water quality in South African catchment","Quantum kernel classifier wins over variational net on E. coli data","75% accuracy: quantum SVM trumps quantum neural network","Support vector quantum classifier defeats neural net on water samples"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000311,"raw_usage":{"total_tokens":1753,"prompt_tokens":909,"completion_tokens":844,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":525,"completion_tokens_details":{"reasoning_tokens":760}},"tokens_in":525,"tokens_out":844,"duration_ms":7014,"temperature":1.0,"reasoning_tokens":760,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T11:27:47.992266+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run leave-one-out cross-validation on the original 32 measurements, keeping the three acceptable samples intact, and report accuracy for each fold; if the QSVC's average held-out accuracy is at or near 50%, or if it never predicts the acceptable class, the claimed QSVC advantage over the QNN is not established.","supporting_citations":[{"cited_title":"Springer, ??? (2018)","cited_arxiv_id":null,"evidence_quote":"Supplies the supervised-learning-with-quantum-computers background that defines the QML task and notation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the quantum machine learning formalism, including feature maps and quantum kernels, used to implement QSVC and QNN."},{"cited_title":"Journal of Computers and Intelligent Systems 3(1), 01–15 (2025)","cited_arxiv_id":null,"evidence_quote":"Reports QSVC and QNN applications to crop yields from water and soil data, against which the paper positions its own domain novelty."},{"cited_title":"Computers 13(8), 191 (2024)","cited_arxiv_id":null,"evidence_quote":"Reports a hybrid quantum deep-learning rice-yield forecast, another QML environmental application used to argue the application space is active."}],"review_version":1}