{"id":"7019f1a3-1a30-4f97-ae17-5b5009ab45c3","arxiv_id":"2601.11423","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"QSMOTE improves non-linear classifiers on imbalanced industrial datasets, and random forests/SVMs tolerate quantum-inspired noise better than linear or naive-Bayes models.","lead":"The paper benchmarks an oversampling method that uses a quantum-style similarity test, called QSMOTE, for imbalanced fault-classification data, and then stresses the method with six noise models. For practitioners, it suggests random forests and SVMs keep working under noisy oversampling while linear and naive-Bayes models degrade, and it proposes this as a baseline for noise-resilient industrial AI.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Algorithm 1's qubit count n = log2(len(c)) + 3 is non-integer for CWRUBD (9 features) and other datasets, making the described swap-test circuit unconstructible; QSMOTE results for these datasets are not reproducible from the stated method.","rationale":"Good-faith reading of the paper shows a carefully structured empirical study with many tables comparing before/after QSMOTE across datasets and classifiers, and a noise robustness analysis. The central claim is that QSMOTE (a quantum-inspired oversampling method using a compact swap test to compute angles) reduces class imbalance and improves non-linear classifier performance, and that injecting noise into the similarity estimation propagates to downstream accuracy. For this claim to hold, the QSMOTE algorithm must be executable on the actual feature matrices.\n\nThe weakest point is in Algorithm 1, which sets the number of qubits to n = log2(len(c)) + 3. This is only an integer if len(c) is a power of two. The paper explicitly states CWRUBD uses nine time-domain features; IFDD uses six sensor readings; SPID is reduced to 32 PCA components (power of two). Thus for at least two of the four datasets, the qubit count is non-integer. A quantum circuit cannot have a fractional number of qubits, and the amplitude-encoding step of a vector of length 2·len(c) requires the number of amplitude qubits to be an integer power-of-two dimension. Without a documented padding or alternative encoding, the circuit cannot be built, and the reported balanced datasets cannot be generated as described. The reader's weakest_assumption identifies exactly this issue; I agree.\n\nThis concern is load-bearing because the primary empirical results (Tables 1-4) all depend on the QSMOTE-generated synthetic samples. If the swap test is not actually run, the results could be obtained by some other interpolation method, but then the paper's methodological contribution (quantum-inspired similarity) is misrepresented. The issue also propagates to the noise analysis: if the base QSMOTE is undefined, the 'noisy QSMOTE' under the six channels is also undefined.\n\nThe concern is not that the conclusions are necessarily false — oversampling generally helps tree-based models — but that the reported numbers are not reproducible from the stated algorithm. This is an internal inconsistency, not a disagreement with consensus, and it is fixable by documenting the padding or switching to a classical angle computation. Hence the appropriate verdict remains CONDITIONAL, not REJECT: the authors can address it with a clarifying implementation note or a corrected algorithm. No further adversarial point is needed; the manuscript contains no code, so this check is the minimum to restore confidence.","tokens_in":32926,"tokens_out":7804,"duration_ms":79065,"concrete_test":"Run the code in Algorithm 1 exactly as written on a 9-dimensional vector (CWRUBD features) using a quantum simulator such as Qiskit's AerSimulator. First, check whether QuantumCircuit(n) accepts the float n=log2(9)+3≈6.17; if it raises a TypeError or requires an integer, the described method cannot execute. Then rerun the QSMOTE pipeline for CWRUBD using either (a) the same code with ceil(log2(len(c)))+3 qubits and zero-padding to the next power of two, or (b) a classical inner-product computation of the angle, and compare the resulting accuracies with Table 2. If the accuracies differ materially, the reported numbers depend on an undocumented implementation choice.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 3.3, Algorithm 1 line 4 sets the number of circuit qubits to n = log2(len(c)) + 3. A quantum circuit requires an integer qubit count. For CWRUBD, Section 4.1 states each segment is summarized by nine time-domain features, so len(c)=9 and n≈6.17; the same issue affects IFDD (six sensor features) and likely EFDD. The amplitude encoding in line 12 initializes a concatenated vector ψ of length 2·len(c) onto qubits 2..n-1, which requires (n-2) qubits to encode 2d amplitudes. This is only possible if d is a power of two; otherwise 2d is not a power of two and cannot be amplitude-encoded without padding. The paper does not describe any padding, rounding, or alternative encoding for non-power-of-two feature counts (e.g., no PCA-to-power-of-two step is mentioned for CWRUBD; the preprocessing section only lists the nine features). Consequently, the swap-test circuit cannot be constructed as written, so the synthetic samples underlying Tables 1-4 and Figures 3-7 for CWRUBD, IFDD, and EFDD cannot be generated by the described method. This directly undermines the central claim that QSMOTE improves classifier performance — the reported results would have to come from an undocumented modification of Algorithm 1.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper presents a benchmark study of Quantum SMOTE (QSMOTE) implemented via a compact swap test, applied to four industrial condition-monitoring datasets (SPID, CWRUBD, EFDD, IFDD) and evaluated with five classical classifiers (LR, RF, SVM, DT, NB) before and after oversampling. The authors also report a robustness analysis under six quantum-inspired noise channels and claim that QSMOTE substantially improves non-linear classifiers, with gains up to roughly 170% on EFDD and near-perfect accuracy on IFDD. The manuscript's main contributions are framed as the first joint evaluation of imbalance mitigation and quantum-noise robustness for industrial fault classification.","tokens_in":33302,"tokens_out":4410,"duration_ms":35928,"significance":"If correct, the paper would provide a useful empirical reference for practitioners combining quantum-inspired oversampling with classical classifiers in industrial diagnostics. The strengths are the breadth of datasets and classifiers, the internally consistent before/after tables in Tables 1–4, and the explicit statement of the main free parameters (κ, K, noise probability grid). I agree with the reader that the tables are internally consistent as comparisons. However, two load-bearing details are missing: the circuit in Algorithm 1 cannot be constructed as written for the datasets used, and the noise-injection protocol is not operationally defined. These gaps prevent the reported results from being reproduced from the described method, so the central claims are not yet established.","major_comments":[{"comment":"The derivation of the overlap estimator is asserted rather than shown. The text writes P0 as a single norm-squared expression, then jumps to Eq. (8), 'after simplifying,' with no intermediate algebra. The notation also mixes state vectors |C⟩,|M⟩ with the scalar normalization constants C and M; Eq. (3) defines |ψ⟩ with states |C⟩,|M⟩ but Eq. (4) and onward appear to treat C and M as scalars. Since the compact swap test is the foundation of the quantum-inspired angle α(x,c) used in Algorithm 2, a correct and complete derivation — or an explicit citation to a source where this relation is proved — is required before the QSMOTE geometry can be considered well-defined.","section":"Section 3.2, Eqs. (7)–(9)"},{"comment":"The conclusion states 'DT on IFDD improving from 0.7393 to 0.9006,' but Table 4 reports IFDD DT accuracy before QSMOTE as 0.4685 and after as 0.9006. The value 0.7393 does not appear in Table 4 and appears to be taken from an unrelated metric or table. This inconsistency suggests a copy‑paste error in a central summary statement and should be corrected.","section":"Section 5, conclusion"}],"minor_comments":[{"comment":"The abstract refers to 'Linear Regression (LR)' while the method section and tables use logistic regression (LR). The model is logistic regression, so the terminology should be corrected.","section":"Abstract"},{"comment":"The single-qubit rotation u(θ, π, 0) is not defined. If this is a Qiskit-style U3 gate, it should be stated explicitly, or the parameters should be named consistently with the circuit diagram.","section":"Algorithm 1, line 8"},{"comment":"The EFDD and IFDD descriptions say each has 1,000 samples but do not state the number of features or the exact class counts before pre-processing. Since Algorithm 1's qubit count depends on the feature dimension, these counts should be reported.","section":"Section 4.1"},{"comment":"Tables A1–A16 are not cited anywhere in the main text, and the appendix is mentioned only as 'Supplementary information.' These additional results should either be integrated into the main narrative or removed, and the caption of Figs. A1/A2 has mismatched subfigure labels (the caption lists (b)–(f) but the figure appears to show different panels).","section":"Appendix A"}],"recommendation":"major_revision","confidential_remarks":"The paper would be a reasonable empirical study for an applied venue if the algorithmic gaps are fixed. My main scope concern is that the quantum component is largely cosmetic: once the qubit-count issue is resolved, QSMOTE reduces to a classical interpolation with a quantum-inspired angle, and the noise analysis is applied to classical classifiers rather than to a quantum pipeline. The authors may want to position the paper more honestly as a benchmark of a quantum-inspired, classically simulated oversampling heuristic. The absence of code or formal verification also makes verification harder, though the fixed random seeds and explicit hyperparameters are a positive step."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a plausible empirical benchmark linking QSMOTE oversampling to quantum-noise robustness across four industrial datasets. The genuinely new part is the combination: nobody else has run QSMOTE plus six noise channels across these datasets and five classifiers. The before/after tables are internally consistent, and the claimed robustness hierarchy (RF/SVM resilient, NB/LR fragile) is a reasonable, useful empirical observation. That part is worth taking seriously.\n\nThe soft spots are real, and one is load-bearing. Algorithm 1 line 4 sets n = log2(len(c)) + 3. For CWRUBD (nine features) and IFDD (six sensors) that is not an integer, and amplitude encoding of a 2d-length vector requires a power-of-two number of amplitudes. The paper never mentions padding or rounding, so the described compact-swap-test circuit cannot be built for those datasets. The reported QSMOTE-generated samples must have come from some undocumented modification. That does not automatically falsify the empirical claims—silent padding would be easy—but the paper as written is not reproducible, and the central comparison rests on an unstated hack. This is fixable: state the padding rule, or use ceil(log2(len(c))) plus explicit zero-padding.\n\nOther issues are smaller but worth naming. The QSMOTE hyperparameters κ and K are never reported, which makes the exact procedure even harder to reproduce. There is no direct comparison against classical SMOTE, so the \"quantum\" oversampling advantage is not actually demonstrated—only QSMOTE before/after. The noise-injection procedure is described at the Kraus-operator level but not mapped to a circuit, which matters because the paper claims noise is injected into the similarity estimation itself. And the conclusion contains a factual slip: it says DT on IFDD improved from 0.7393 to 0.9006, but Table 4 shows the pre-QSMOTE DT accuracy as 0.4685. Minor, but sloppy.\n\nWho is this for? Practitioners who want a first map of how classical models behave under noise-perturbed quantum-inspired oversampling. The idea is worth testing properly, and the empirical pattern is plausible. But the method is not fully specified and no code or data artifacts are shipped to compensate.\n\nI would send this to peer review, because the combination is novel and the flaws are fixable. Referees should be specifically asked to check the qubit-count issue and to require either a corrected algorithm with explicit padding, or code plus a classical SMOTE baseline. As it stands, the paper is a promising draft, not a finished result.","headline":"A useful empirical benchmark idea undercut by a non-integer qubit count in Algorithm 1 that makes the core QSMOTE results unreproducible without an undocumented workaround.","tokens_in":33752,"tokens_out":3036,"would_cite":false,"duration_ms":33856,"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":"QSMOTE is claimed to fix class imbalance in industrial fault data, boosting random forests and SVMs by up to 170% on engine-failure data and to over 0.99 accuracy on industrial-fault data, while linear and naive Bayes models often degrade.","keywords":["quantum SMOTE","class imbalance","oversampling","fault diagnosis","condition monitoring","quantum noise","noise robustness","machine learning"],"falsifier":"Inspect the experiment code or rerun the pipeline for a dataset with a non-power-of-two feature count, such as CWRUBD's nine time-domain features. If Algorithm 1 is applied as written, constructing a circuit with n = log2(9) + 3 qubits (a fractional number) should fail in any standard quantum circuit simulator; if the code silently pads, truncates, or switches to a classical inner product, then the QSMOTE samples were not generated by the amplitude-encoded compact swap test, and the reported accuracy gains are attributable to whatever substitute was actually used. A reader could also compare t","tokens_in":32834,"feed_emoji":"⚙️","tokens_out":10130,"duration_ms":85335,"temperature":0.7,"pith_summary":"QSMOTE is claimed to solve class imbalance in industrial and energy fault-detection datasets by generating synthetic minority samples whose step length is governed by a quantum-inspired similarity angle. Across four datasets, the paper reports consistent large gains for non-linear classifiers—random forests, decision trees, and SVMs—with relative improvements up to about 170% on the Engine Failure Detection Dataset and near-perfect accuracy (over 0.99) on the Industrial Fault Detection Dataset. The same results show that linear and probabilistic models, particularly logistic regression and naive Bayes, often lose accuracy after oversampling, which the paper attributes to interpolation-induced boundary distortion in overlapping feature spaces. In a second stage, the paper injects six quantum-noise channels into the similarity-estimation step and claims that ensemble and margin-based models retain most of their accuracy even at maximum noise, while naive Bayes and decision trees degrade substantially, with depolarizing noise the most disruptive. If these claims hold, they give practitioners a clear baseline: use QSMOTE with an ensemble or SVM for imbalanced, noisy industrial monitoring, and avoid simple linear or probabilistic models.","feed_headline":"Lift fault-detection accuracy to 0.99 with quantum-inspired sampling","feed_subtitle":"Non-linear models gain up to 170% on engine-failure data; linear and naive Bayes models often degrade.","key_machinery":"The mechanism is the compact swap test used to estimate the quantum overlap between a minority sample and its cluster centroid. The sample and centroid are amplitude-encoded as quantum states; the circuit measures the ancilla in a way that encodes their inner product, which is converted to a 'quantum-inspired angle' by arccos after clipping to [-1,1]. A uniform step length is then sampled from [0, angle/split_factor], and the synthetic sample is placed at x + step * (centroid - x)/||centroid - x||. The angle therefore directly controls how far synthetic points move: a poorly aligned sample (large angle) moves farther toward its centroid, while an aligned sample stays close to x. The noise an","core_discovery":"The paper's central claim is that QSMOTE—which generates a synthetic minority sample by moving the point toward its K-means cluster centroid by a distance drawn uniformly from zero to a 'quantum-inspired angle' between the point and the centroid divided by a split factor—substantially improves non-linear classifiers on imbalanced industrial fault data. The reported effect is model- and dataset-dependent: on EFDD, random forest accuracy rises from 0.4923 to 0.9126 and decision tree from 0.2979 to 0.8228; on IFDD, random forest reaches 0.9919 and decision tree 0.9006; on SPID, SVM rises from 0.8057 to 0.8893. Logistic regression and naive Bayes often fall, with LR on EFDD dropping from 0.5337","pith_inferences":["A classical variant may capture most of the effect: if the gain comes mainly from the angle-scaled step toward cluster centroids, replacing the swap-test angle with a classical cosine similarity between x and its centroid would be a cheap, falsifiable proxy to test.","The optional convex clamp mentioned in the paper is not used in Algorithm 2; enabling it would force synthetic points onto the segment from x to its centroid and is a natural extension to see whether the reported LR/NB degradation shrinks.","Noise injected into the similarity estimator is a distinct failure mode from noise on raw sensor features; comparing the two would show whether quantum-inspired perturbation models are the right stress test for real industrial sensor corruption.","Because the swap-test estimate is based on a fixed shot count (1000), sampling noise in the angle itself is a source of stochasticity; increasing shots should tighten angle estimates and could reduce variance in the reported gains, a testable prediction."],"forward_implications":["For imbalanced industrial and energy monitoring data, QSMOTE plus a random forest or SVM is a strong combination: reported accuracies exceed 0.99 on IFDD and 0.98 on CWRUBD.","Model-family guidance emerges: non-linear, ensemble, and margin-based classifiers benefit from QSMOTE; logistic regression and naive Bayes often degrade, so resampling choices must be model-aware.","The noise results give a simple selection rule for noisy environments: random forests retain near-baseline accuracy at maximum noise, while naive Bayes and logistic regression should be avoided when quantum-inspired perturbations are possible.","Depolarizing noise being the most disruptive channel and phase-damping the least suggests where to direct noise-mitigation effort in quantum-inspired resampling pipelines.","The reported 170% relative gain on EFDD provides a concrete performance target for validating future balancing methods against this baseline."],"fun_headline_variants":["Quantum SMOTE lifts fault accuracy to 0.99","QSMOTE boosts non-linear fault models up to 170%","Fault detection accuracy hits 0.99 with quantum SMOTE","Quantum-inspired oversampling: 99% fault accuracy","Fault classification: QSMOTE aids non-linear, hurts linear"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The paper's load-bearing premise is that the compact swap test in Algorithm 1 can actually be executed on the datasets' feature vectors; the qubit count is set to log2(len(c)) + 3, which is not a whole number for feature vectors whose length is not a power of two (CWRUBD has nine features), and no padding or alternative amplitude-encoding step is described—if the circuit cannot be constructed, the reported synthetic samples and all downstream accuracy figures are not produced","fun_headline_variants_meta":{"raw":{"variants":["Quantum SMOTE lifts fault accuracy to 0.99","QSMOTE boosts non-linear fault models up to 170%","Fault detection accuracy hits 0.99 with quantum SMOTE","Quantum-inspired oversampling: 99% fault accuracy","Fault classification: QSMOTE aids non-linear, hurts linear"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000407,"raw_usage":{"total_tokens":2017,"prompt_tokens":879,"completion_tokens":1138,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":623,"completion_tokens_details":{"reasoning_tokens":1066}},"tokens_in":623,"tokens_out":1138,"duration_ms":11947,"temperature":1.0,"reasoning_tokens":1066,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T10:00:47.385402+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the experiment code or rerun the pipeline for a dataset with a non-power-of-two feature count, such as CWRUBD's nine time-domain features. If Algorithm 1 is applied as written, constructing a circuit with n = log2(9) + 3 qubits (a fractional number) should fail in any standard quantum circuit simulator; if the code silently pads, truncates, or switches to a classical inner product, then the QSMOTE samples were not generated by the amplitude-encoded compact swap test, and the reported accuracy gains are attributable to whatever substitute was actually used. A reader could also compare t","supporting_citations":[],"review_version":1}