In the paper Efficient Bitcoin Address Classification Using Quantum-Inspired Feature Selection, feature selection is formulated as QUBO and combined with classical Bitcoin address classifiers. Published data show that switching from 69 to 23 features selected by simulated annealing reduces Random Forest training from 305 min 5.7 s to 212 min 23.2 s while changing the Mixer class F1 from 0.92 to 0.91. [file:1] A secondary recalculation yields a savings of 30.39% and a speedup of about 1.44x; these numbers do not prove a quantum advantage. The paper links studies of summary transaction features, distribution moments, and QUBO selection, identifies metric ambiguities, and proposes a controlled experiment with an equal number of features, nested validation, and full cost accounting. The central result is a map of testable explanations that allows one to distinguish the benefit of a mathematical formulation from the benefit of a quantum device.
Keywords: quantum annealing; simulated annealing; QUBO; Bitcoin; address classification; feature selection; reproducibility; quantum advantage.
1. Paradox: Quantum idea, classical winner
Could the practical utility of quantum technologies be improved by discovering that a conventional computer performs better in a given experiment? The work by Ming-Fong Xie, Yen-Jui Chang, Chien-Lung Lin, Ching-Ray Chang, and Shih-Wei Liao offers just such a rationale for research. The authors transform Bitcoin address feature selection into a binary optimization, test several solvers, and obtain the most convincing speed-performance tradeoff using classical simulated annealing (SA). [file:1]
This article does not focus on hacking Bitcoin or recovering a private key. Rather, it examines the chain “transaction history → features → feature selection → classifier.” The central question is: does a quantum solver speed up a useful problem with comparable performance, or is the observed gain due to a reduction in the classical model’s input?
The working conclusion: the published data supports the usefulness of a compact feature set for Random Forest, but does not establish a quantum advantage. A more interesting next step is to separate the contribution of the mathematical setup from that of the physical device.
Our contribution is a secondary quantitative analysis, a consistency audit of the results, and a decisive comparison project. No retraining of the models, access to the quantum device, or independent replay of the dataset were performed. The training example below is accurately calculated but is not a Bitcoin experiment.
2. Research method and limits of evidence
The methodological basis is a provided construction kit consisting of 12 sections: question formulation, mechanism, history, apparatus, mathematical model, reliability, fair comparison, reproducibility, cryptography, related technologies, practical value, and further research. It explicitly recommends selecting questions rather than answering them all at once. For this topic, 18 general and 6 specialized questions were selected; their operationalization is provided in Section 13. [file:2]
The main material is all 21 pages of the provided PDF, including Tables 1–10, figure captions, and references. The PDF version is arXiv:2411.15425v1, dated November 23, 2024. The verified arXiv file shows the same version. The page number in the arXiv metadata differs from the local file; the page numbers below refer specifically to the provided PDF. The analysis is not intended to be exhaustive of the state of the field in 2026. [file:1] [web:18]
- Level I: Directly read methods and numbers of the main PDF.
- Level II: independent arithmetic recalculation of published numbers and logical analysis of formulas.
- Level III: Reviewed related work cards, the 1QBit page, and the repository README. These documents confirm the information provided therein, but do not replace an audit of the full code and all source data.
- Level IV: Proposed hypotheses and protocols. These have not yet been tested and are not presented as results.
For Toyoda et al., the method was linked to by reference [25] and the conclusion of the main PDF; the full source has not been independently verified. For Lin et al., the arXiv abstract was directly verified. Therefore, details missing from the available evidence are not attributed to these publications below. [file:1] [web:5]
3. How several works form one study
| Work and verification status | What does she add? | What is the next question that arises? |
|---|---|---|
| Toyoda, Ohtsuki, Mathiopoulos, 2018; [25] main PDF, indirect verification | Classification of Bitcoin services through a summary description of transaction history; the authors of the main paper associate basic statistical features with this work. [file:1] | Are simple sums and frequencies sufficient to distinguish behavior? |
| Lin, Wu, Hsu, Tu, Liao, 2019; ICBC; abstract reviewed | High-order moments of transaction time, expressed as block height. For LightGBM, Micro-F1 is reported at 87% and Macro-F1 at 86%. [web:5] | Is it possible to preserve information by removing redundant features? |
| Milne, Rounds, Goddard, 2017; white paper 1QBit | QUBO-based selection of influential and weakly dependent features for a credit scoring problem; potential application of quantum annealing. [web:15] | Is the production portable to Bitcoin and does it require a quantum solver? |
| Kirkpatrick, Gelatt, Vecchi, 1983; [16] main PDF | Classical simulated annealing is a mandatory control line for QUBO, not quantum hardware. [file:1] | Does QA outperform strong classical heuristics? |
| Hauke et al.; preprint 2019, journal publication 2020 | A review of methods, implementations, and obstacles to scaling quantum annealing. [web:6] [web:3] | What proportion of the limitations are manifested in the real application chain? |
| Sie et al., 2024; main preprint | Summary of Bitcoin + QUBO + SA / QA / hybrid BQM + classic classifiers. [file:1] | What exactly explains the gain, and will it be maintained across new services and periods? |
This story has two converging branches: the development of transaction representation and the development of feature selection optimization. Their combination doesn’t prove a physical speedup. It creates a common experimental interface on which different solvers can be compared.
You can’t subtract Lin et al.’s 87% Micro-F1 from Sie et al.’s 91% F1 for the Mixer class and claim a four-percentage-point improvement. Micro-F1 aggregates decisions across objects, Macro-F1 averages equally weighted class F1s, and Mixer F1 is specific to a single class. Additionally, the data, partitions, and tuning mode must match. [web:5] [file:1]
4. What exactly was measured
The authors report processing 544,283 blocks and 345,882,038 transactions, four months of data acquisition, and a limit of the first 1,000 transactions per address. The hardware configuration includes an Intel Core i9-12900K, 64 GB of RAM, a GTX 1060 6 GB, and a 4 TB SSD. The presence of a GPU in the configuration alone does not indicate that all classifiers used a GPU. [file:1]
| Class | Marked addresses, table 1 | Interactive Transactions, table. 1 |
|---|---|---|
| Exchange | 163,537 | 2,582,337 |
| Faucet | 14,720 | 87 245 |
| Gambling | 78,698 | 170,815 |
| Market | 102 129 | 181,337 |
| Mixer | 289,006 | 234,449 |
| Pool | 46,586 | 203,590 |
| Total | 694 676 | 3,459,773 |
The source of the annotation is WalletExplorer. The same table also lists 1,048,575 “Interactive Addresses” separately; this should not be automatically considered the number of independent annotated examples. A transaction can interact with multiple addresses, so interaction aggregates cannot be considered unique transactions without defining a methodology. [file:1]
The Mixer share of the labeled addresses is 289,006 / 694,676 ≈ 41.60%—this is a recalculation of Table 1. The share of mixers in a random address stream is not shown. The authors use 69 features and seven classifier families: Logistic Regression, AdaBoost, Random Forest, XGBoost, LightGBM, SVM, and Neural Network. Some SVM and Neural Network rows represent only 10% of the data. [file:1]
Features include transfer amounts and frequencies, address lifetime, balance statistics, distribution moments, and input/output patterns. However, the promise of a “full history” for feature calculations comes with a cutoff of 1,000 transactions. Without a precise cutoff rule, active address statistics may have a different meaning than the terms lifetime or total imply. This is a testable risk, not a fixed error for all results. [file:1]
5. Where is physics and where is software optimization?
QUBO stands for quadratic unconstrained binary optimization. It’s a mathematical language, not a type of computer. BQM stands for binary quadratic model; the term itself also doesn’t denote quantum computing. In the main paper, SA is implemented using dwave-neal; for QA, the D-Wave Advantage system 6.4 with 5612 working qubits is specified, and for general BQM, the D-Wave Hybrid solver 2.2 is used. The latter should be called a hybrid solver: its name doesn’t reveal the share of QPU work. [file:1]
SA modifies classical binary variables and allows energetically unfavorable transitions with a temperature-dependent probability; with cooling, such transitions become less frequent. This is a thermal heuristic, not physical quantum tunneling. The article uses the umbrella term “quantum-inspired,” but this does not make SA a quantum algorithm. [file:1]
An idealized description of quantum annealing uses a changing Hamiltonian:
H(s) = −A(s) ∑ᵢ σᵢˣ + B(s) [∑ᵢ hᵢ σᵢᶻ + ∑ᵢ<ⱼ Jᵢⱼ σᵢᶻσⱼᶻ], 0 ≤ s ≤ 1.
Initially, the transverse field dominates, then the role of the Ising target increases. Quantum fluctuations allow for dynamics different from thermal SA jumps. However, the physical feasibility of tunneling is not a measure of classification gain. A real device operates at a finite temperature, with limited connectivity and uncertain control parameters; an idealized model does not guarantee finding a global minimum in a practical time. [web:6]
A tightly coupled logic problem may require minor embedding: one logical variable is represented by a chain of physical qubits. These variables are not the logical qubits of a fault-tolerant quantum computer. The number 5612 does not represent 5612 independent fully coupled features, nor does it have the resources to run Shor’s algorithm. The PDF does not provide sufficient information about the chains, their breaks, and calibration to determine the cause of the QA degradation. [file:1] [web:6]
6. Mathematics: What is actually being optimized?
For the binary Mixer vs. Rest problem, we define xⱼ = 1 when the feature j is included, rⱼ = |ρ(y, Dⱼ)|, and cⱼₖ = |ρ(Dⱼ, Dₖ)|. Spearman’s rank correlation is used. The basic PDF is written as:
f(x) = −α ∑ⱼ rⱼxⱼ + (1−α) ∑ⱼ ∑ₖ≠ⱼ cⱼₖxⱼxₖ, xⱼ ∈ {0,1}, 0 ≤ α ≤ 1.
The first term rewards the association with the label, while the second penalizes pairwise dependence. This transfers the influential/independent idea from Milne et al. to Bitcoin classification. The correlation surrogate is minimized, not the Random Forest F1 directly. Therefore, a lower QUBO value does not necessarily yield a better classifier. [file:1] [web:15]
For symmetric correlations, the double sum counts each pair twice. In the xᵀQx convention, the symmetric matrix has Qⱼⱼ = −αrⱼ and Qⱼₖ = (1−α)cⱼₖ for j ≠ k; in upper-triangular storage, the coefficient of the pair is 2(1−α)cⱼₖ. When using −xᵀQx, the signs must be consistently reversed. Coefficients cannot be transferred between libraries without checking the conventions. This is a direct algebraic consequence of the PDF formula.
The text talks about selecting K features, but the given function does not impose ∑xⱼ = K. The size arises from α, correlations, and the found solution. For a controlled comparison, one can propose:
f K (x) = f(x) + λ(∑ⱼ xⱼ − K)².
This is a proposed addition, not a stated result of the original work. λ needs to be justified and verified; too small a penalty allows for a violation of K, while too large a penalty degrades the scale of the coefficients on the device. An alternative is a separate solver with an explicit limit on the set’s cardinality.
Perfect correlation does not imply statistical independence. Zero pairwise correlation can coexist with nonlinear dependence; an important combination of individually weak features may be underestimated. The selection function is a useful engineering hypothesis, not a proven optimal criterion for Random Forest.
A minimal example with an exact answer
We set the artificial estimates r = (0.80; 0.75; 0.40), c₁₂ = 0.90, c₁₃ = 0.10, c₂₃ = 0.15, and α = 0.70. These are training coefficients, not blockchain estimates. A complete enumeration of the eight sets yields:
| Set x₁x₂x₃ | f(x) |
|---|---|
| 101 | -0.780 |
| 011 | -0.715 |
| 111 | -0.675 |
| 100 | -0.560 |
| 110 | -0.545 |
| 010 | -0.525 |
| 001 | -0.280 |
| 000 | 0.000 |
The minimum is reached at 101: the first and third features together are more useful than the two strongest ones individually, since the first pair is heavily penalized for redundancy. Enumeration is needed here as a benchmark: first, the coding correctness and the quality of the heuristics are checked on a small problem, and then scaling is discussed.
7. Recalculating the results: where is the winnings?
The key values below are transferred from PDF tables 3 and 6. The QA and SA sets are not the same size; the same number of features, tuning budget, and repeated measurements are not provided by the published table. [file:1]
| Model | Signs | Selection, with | Education | Precision / Recall | F1 Mixer |
|---|---|---|---|---|---|
| Random Forest, all the signs | 69 | — | 305 min 5.7 sec | 0.87 / 0.97 | 0.92 |
| SA + Random Forest | 23 | 0.0055 | 212 min 23.2 sec | 0.85 / 0.97 | 0.91 |
| QA + Random Forest | 9 | 6,644 | 125 min 27.7 sec | 0.58 / 0.62 | 0.60 |
| Hybrid BQM + Random Forest | 7 | 6,0041 | 183 min 17.5 sec | 0.65 / 0.76 | 0.75 in Table 6; about 0.70 in the text |
Savings = 1 − 12,743.2 / 18,305.7 = 0.303867… ≈ 30.39%.
Training speedup = 18,305.7 / 12,743.2 ≈ 1.4365 times.
Saved 5562.5 s = 92 min 42.5 s.
The abstract states 30.3%. A direct conversion of the printed times yields 30.39%, or 30.4% with the usual rounding to one decimal place. This is a minor rounding/presentation discrepancy, not a refutation of the gain itself. The number of features decreased by 66.67%, but the training time was only approximately 30.39%: the model cost is not necessarily linear in the number of columns. [file:1]
F1 decreased from 0.92 to 0.91: a decrease of 0.01, or one percentage point. The wording “without loss of quality” is too strong without a pre-selected tolerance and confidence interval for the difference. However, such a compromise can be useful if the acceptable loss of quality is agreed upon before the experiment.
The number 6.644 / 0.0055 = 1208 indicates that the published QA stage time is 1208 times longer than the SA time. For Hybrid BQM, the ratio is approximately 1092. These ratios describe specific table times, not the fundamental speed of the QPU relative to the CPU: the overhead accounting boundaries and hardware modes are not well-covered. [file:1]
Other classic models cannot be ignored. XGBoost without selection has an F1 of 0.88 at 70 min 5 s, LightGBM has an F1 of 0.87 at 56 min 1.8 s. They are faster than SA Random Forest, but worse by this metric; they are choices at the quality-cost boundary, not overall champions. Moreover, SA XGBoost in Table 6 drops from an F1 of 0.88 to 0.58: the set useful for Random Forest is not universal. [file:1]
8. Audit: What discrepancies cannot be hidden
The following points pertain to the version provided, not to possible revisions. They require clarification from the authors or independent verification, but do not in themselves prove dishonesty. [file:1]
| Observation | Why is it important? | How to check |
|---|---|---|
| Hybrid BQM RF: F1 = 0.75 in Table 6; about 0.70 in the descriptions of Section 4 and Figure 9 | From the printed P = 0.65 and R = 0.76, the harmonic mean is ≈ 0.7007. However, the average F1 across folds is not equal to the F1 from the average P/R; one recalculation does not determine the correct number. | Request predictions, support, and component-specific metrics for each fold. Do not use 0.75 as a consensus value until verified. |
| QA RF: AUC = 0.88 in Table 6 and 0.93 in Table 7 | The same named model produces different estimates; it is unclear whether the aggregation procedures are the same. | Check scores, positive class, weights and averaging method. |
| SA XGBoost: AUC = 0.89 in Table 6 and 0.98 in Table 7; SA Neural Network: 0.98 vs 0.88 | The discrepancies affect more than one line; the overall AUC rating is unreliable without verification. | Rebuild both tables from the single saved prediction array. |
| The narrative text contains “recall 0.97%” and similar entries | A proportion of 0.97 means 97%, not 0.97%. Confusion changes the meaning a hundredfold. | In this article, metrics are given as shares, percentages are clearly signed. |
| The SA set in Table 4 contains m₂,interval and m₃,interval; the importance table 9 includes m₁,interval | If it is the same set, the importance of the unselected feature requires explanation. | Check the serialized selection mask against the column names of the trained model. |
| The text calls the four transaction types mutually exclusive, but the spent rule overlaps with payback. | The address in the inputs and outputs satisfies both verbal conditions without explicit priority. | Publish the exact order of the rules and tests for return of change. |
Additional consistency check: For a binary test with Mixer fraction π = 0.4160, P = 0.87, and R = 0.97, FP/N = πR(1/P − 1) holds. Then accuracy = 1 − π(1−R) − πR(1/P−1) ≈ 0.927, not 0.95. This is a conditional diagnostic calculation, not a proven error: the table may use a different test composition, averaging, or a different definition of accuracy. But the definitions, support, and confusion matrix are necessary to explain such a difference. [file:1]
Formatting errors don’t negate the potential value of an idea. They reduce the accuracy of possible conclusions and make preserving the original predictions more important than a beautiful final table.
9. What is missing for causal inference?
9.1 Different number of features – mixing of factors
Comparing 23 SA features with 9 QA features and 7 Hybrid BQM features simultaneously varies the solver, dataset size, and input composition. The drop in F1 may be due to excessive compression, not quantum physics. The assumption that “QA is worse due to noise” is as unproven as “QA is faster due to tunneling.” Experiments should fix K and separately vary the solver. [file:1]
9.2. Pre-split selection may reveal test labels
The correlation with y uses labels. If it is calculated on the entire sample before cross-validation, the test portion will influence the selection. The PDF describes a 10-fold CV, but does not provide sufficient evidence that all selection and tuning steps are embedded within the training folds. Therefore, leakage is a risk to be excluded, not an established fact. [file:1]
9.3 Addresses can be dependent
Randomly partitioning addresses can distribute addresses of a single service between training and testing. This way, the model recognizes a familiar service rather than transferring knowledge to an unknown one. Validation requires service-disjoint partitioning and a temporal test with features calculated only from the past. When studying temporal transfer, it is impossible to aggregate the entire future history of an address.
9.4 Markup and behavior change
WalletExplorer labels describe the data source and its rules, not the automatically verified legal status of an address. The snapshot date, the origin of the labels, and the rules for handling ambiguous addresses are required. A model trained on a historical dataset cannot be proven suitable for new services without a separate test. [file:1]
9.5 Uncertainty and reproducibility
A sufficient set of interval-based repeated measures is not provided to interpret an F1 difference of 0.01 as established equivalence. The repository’s README describes the sequence preprocessing.py → summarization.py → notebooks for preparation, SA, and classification. It states a minimum of four days for preprocessing and 4–8 hours for summarization for 200 million transactions. These are independent estimates from the README, not a universal standard or the duration of the main experiment. [web:34]
The GitHub link in the extracted PDF text contains a space, but the PDF annotation also contains the correct URL with an underscore; this is the one that was accessed. The presence of a README does not mean that the code, dependencies, dataset availability, license, or reproducibility of the results have been fully verified.
10. Three hypotheses that can actually be refuted
The threshold values below are proposed for a future protocol. These are not published results or generally accepted standards.
| Hypothesis | Operational criterion | What refutes / does not confirm |
|---|---|---|
| H1. Feature compression maintains almost acceptable quality and saves overall time. | For SA RF versus RF-69: the lower limit of the 95% confidence interval ΔF1 is above −0.01; the upper limit of the interval of the ratio of total times is below 0.80 in a pre-specified operating scenario. | A significant drop in F1 of more than 0.01 or the absence of savings refutes the corresponding part. An interval crossing the threshold signifies uncertainty, not proven equivalence. |
| H2. For the same K, QA provides practical advantages over strong classical solvers. | Given fixed K, data, and budgets, QA must achieve a predetermined quality and success probability faster than the best classical line, taking into account preparation, embedding, and processing. | The lack of a consistent gain on the test family does not confirm an advantage in this mode; it does not disprove all possible applications of QA. |
| H3. Predictive ability carries over to unknown services and future periods. | Pre-proposed threshold: Mixer F1 ≥ 0.85 on service-disjoint temporary holdout and acceptable false alarm level specified by the system user. | Falling below the threshold, especially after excluding leakages, rejects the transferability in the chosen scenario, but not the usefulness of the model on the original set. |
The available tables can be used to test the arithmetic of learning savings, but H1, H2, or H3 cannot be fully accepted: the required intervals equal to K and an independent transfer test are not available.
11. The Crucial Experiment: From Replication to Fair Comparison
- Capture a blockchain snapshot, elevation boundaries, exchange rates, labels, date, and input file hashes. Release a dictionary of 69 features with units and spent/payback resolution rules.
- Set external tests: random address as a control, service-disjoint as a transfer check, and time as a prediction check. Choose fold sizes based on the number of independent services, rather than automatically assigning ten.
- Within each outer train, perform cleaning, imputation, correlations, α and K fitting, classifier selection, and tuning. The outer test should not be used for either method selection or decision thresholding.
- Compare RF with all features; random subsets as distribution; Spearman ranking; mRMR or RFE; QUBO with SA; QA; Hybrid BQM. Add XGBoost and LightGBM as alternative chains. Set tuning budgets identically or explicitly report differences.
- Check K ∈ {7, 9, 23}; compare the solvers of a single QUBO for each K. Allow each method to separately find K in the inner loop: this is a different issue; its results should not be mixed with those of a fixed K.
- For small problems, obtain the exact optimum by brute force or an exact solver and check the QUBO–Ising coefficients. For large problems, compare with the best known value without declaring it a proven minimum.
- For example, pre-assign 30 independent solver seeds to a fixed input and a balanced tuning budget. Seed repetitions measure the solver’s randomness rather than increasing the number of independent addresses or services.
- For QA, preserve embedding, chain lengths and strengths, break rate, num_reads, annealing_time, gauge transforms, device parameters, date, and solver ID. For hybrid, disclose available telemetry and do not attribute the entire result to QPU.
- Perform ablations: without USD attributes; without moments; without transaction patterns; the first 100, 1000, and all available past history. Verify that currency attributes do not simply encode the period and service composition.
- Save out-of-fold probabilities, labels, support, feature masks, and stage durations. Generate tables automatically from a single set of artifacts.
What to measure
- Quality: per-class precision/recall/F1, Macro-F1, Micro-F1, ROC-AUC and PR-AUC for Mixer, error matrix and selected operating threshold.
- Optimization: energy, gap to the proven or best known solution, probability of achieving the given goal and stability of the selected features.
- Resources: CPU/GPU/QPU wall-clock, memory, access cost, and energy if measured; don’t guess at the unknown.
- Uncertainty: paired differences on single external tests, intervals accounting for grouping by service and time dependence. A simple standard deviation of ten overlapping CV folds does not independently provide evidence of significance.
Full cost instead of a successful table row
T total = T data + T features + T corr + T tuning + T selection + T train + T post .
For cloud QA, show queue, embedding, programming, annealing, readout, and data transfer separately. Two reports are needed: the computational cost of the method without an external queue and the user latency with a queue. Don’t add the durations of different stages together if the provider has already included one within another.
It’s also important to determine in advance how much retraining a single selection will require: a one-time experiment and repeated retraining have different depreciation. If history collection and feature extraction are shared and dominant, the gain from training alone yields significantly lower relative savings across the entire chain.
N = ceil[ln(1−p*) / ln(1−p)], TTS = T setup + N · t run .
Here, p is the empirical probability of a single independent run reaching a given energy threshold, and p* is the target success probability, for example, 0.99. The formula is applicable for 0 < p < 1; for p = 0, there is no final estimate, and for p = 1, a single run is sufficient. For correlated readouts, independence must be verified. This is the definition of the proposed protocol, not the measured TTS from the PDF. The energy TTS does not replace downstream F1.
12. Bitcoin, Privacy, and the Boundary of Cryptanalysis
The model under study receives public transaction history and service labels. It produces a predicted address class, not a private key, owner identity, or proof of a crime. The PDF generally does not demonstrate solving the secp256k1 discrete logarithm, counterfeiting ECDSA/Schnorr, accelerating mining, or running Shor’s algorithm. Therefore, it cannot be used to infer a timeframe for “hacking Bitcoin with a quantum computer.” [file:1]
The problem relates to behavioral analysis and privacy, but is not signature cryptanalysis. For this problem, questions about the moment of public key disclosure, Shor’s logical resources, and post-quantum migration do not define the experiment. Questions of quantum key distribution, sensors, clock accuracy, and quantum simulation are also irrelevant to this work. These are deliberate boundaries, not gaps to be filled with unrelated promises. [file:2]
A false classification can have consequences for the user. Even if the original sample has a precision of 0.85, this represents the proportion of correct responses in a specific test, not the probability of the address being at fault. To illustrate the base rate transfer, let’s assume sensitivity = 0.97, FPR = 0.01, and Mixer prevalence = 0.001:
PPV = sensitivity π / [sensitivity π + FPR (1−π)]
= 0.00097 / (0.00097 + 0.00999) ≈ 8.85%.
All three numbers in this example are artificial; the FPR of 1% is not extracted from the PDF. This example demonstrates why even high recall is no substitute for measuring precision on a real flow. The Mixer marker alone does not establish the illegality of a particular operation.
A practical system should preserve the label’s origin, model version, and signal causes, provide for human verification, and correct misattributions. These are proposed requirements for use, not a description of a service already implemented by the authors.
13. Matrix of questions from the research constructor
Eighteen questions were selected from the general framework, and six specialized ones were added. The wording was shortened and adapted, and not presented as literal quotes or standards. The section numbers in the last column refer to this article. [file:2]
| Group | Question to be tested | Response / Check Status | Where revealed |
|---|---|---|---|
| General: problem | What is being reduced? | Time of classical training after selection; total time still to be measured. | §1, 7, 11 |
| General: paradox | Why is classic SA stronger than QA in the table? | The set size, criterion, and setting are possible; there is no causal answer yet. | §7, 9 |
| General: object | What is the unit of analysis? | Signs of address history; not a separate signature or key. | §4, 12 |
| General: Falsifiability | What refutes usefulness? | Unacceptable reduction in F1 or lack of full chain savings. | §10 |
| General: boundaries | What is not being researched? | Signature hacking, mining, sensors, networks, and PQ migration. | §12 |
| General: mechanism | Where is quantum physics? | In the QA device, but not in SA and not in QUBO itself. | §5 |
| General: minimal example | Is it possible to check the model manually? | Eight subsets of three features provide an accurate standard. | §6 |
| General: history | How are publications combined? | Summary → Highlights → QUBO → Solvers → Classification. | §3 |
| General: Alternatives | What approaches compete? | SA, QA, hybrid, ranking, RFE/mRMR and full models. | §11 |
| General: hardware resource | What does 5612 qubits mean? | The physical resource of the device, not the number of fault-tolerant logical qubits. | §5 |
| General: Preparing the entrance | Is the acquisition of features included in time? | In the published savings of training – not as a complete end-to-end report. | §7, 11 |
| General: Model Limitations | Is F1 minimized? | No, the correlation surrogate is minimized; K is clearly not fixed. | §6 |
| Overall: Fair Comparison | Are K and budgets the same? | In the main table K differs; a controlled test is needed. | §7, 11 |
| Overall: Level of Evidence | What is done in this article? | Audit and recalculation; no new hardware tests. | §2 |
| General: uncertainty | Is there a proven equivalence of F1? | There are not sufficient intervals to conclude that there is no deterioration. | §9, 10 |
| General: reproducibility | Is the code link sufficient? | README confirmed; full reproducibility not verified. | §9 |
| General: practical value | What is the cost of a mistake? | False attribution and the dependence of precision on base frequency. | §12 |
| General: Next Research | How to distinguish explanations? | Equal K, nested partitioning, service/time holdout and ablations. | §11 |
| Profile: Mathematical model | How is QUBO related to Ising? | Through binary variables and the transformation xᵢ=(1+sᵢ)/2; the coefficients are checked on a small standard. | §5–6 |
| Profile: division of labor | What does a QPU do and what does a CPU do? | QPU solves the selection problem in QA; features and classifiers are classical; hybrid requires telemetry. | §5, 11 |
| Profile: Connectedness | What is the cost of embedding? | Chains and breaks are important; there is insufficient data for causal inference. | §5, 11 |
| Profile: reliability | How to account for noise and randomness? | Restarts, target energy, intervals and device parameters. | §11 |
| Profile: Scaling | Is it proven that it accelerates as n increases? | No; families of real problems and strong classical lines are needed. | §10–11 |
| Profile: Threat Model | What data does the analyzer have and what does it learn? | Public transactions and labels → likely service class; not private keys. | §12 |
14. What does this change in our understanding of quantum progress?
The works reviewed are compatible at the research program level: a better transaction description increases the information content of the input, and selection attempts to reduce redundancy. However, the usefulness of a QUBO setup does not determine the usefulness of a specific quantum processor. A negative or ambiguous result from a device helps to localize the next experiment, if differences in K, quality, and timing are not obscured. [file:1] [web:5] [web:15]
Instead of promising a breakthrough date, four measurable milestones make sense:
- Independently reproduce the compact SA set with clean test and full time.
- Set the same quality of QUBO solutions for different solvers at equal K.
- Demonstrate consistent cost gains with comparable downstream F1 across multiple independent tasks.
- Test the migration to new services and periods, and then assess operating costs and the consequences of errors.
The bottom line based on the available data: SA Random Forest offers an observable tradeoff between training time and quality; the reported savings of approximately 30% do not equate to quantum speedup. It remains unknown whether QA can benefit from equalizing conditions and eliminating methodological uncertainties. It is precisely this unknown that transforms this review into a research agenda rather than a sales pitch.
15. Appendix: Short code for checking arithmetic
The code uses only the standard Python library. It reproduces the secondary recalculation and QUBO training, but does not train the classifier or access D-Wave.
from itertools import product
baseline = 305 * 60 + 5.7
selected = 212 * 60 + 23.2
print("Экономия, %:", 100 * (1 - selected / baseline))
print("Ускорение:", baseline / selected)
print("Сэкономлено секунд:", baseline - selected)
p, recall = 0.65, 0.76
print("F1 из напечатанных P/R:", 2 * p * recall / (p + recall))
alpha = 0.7
relevance = [0.8, 0.75, 0.4]
redundancy = {(0, 1): 0.9, (0, 2): 0.1, (1, 2): 0.15}
def energy(x):
influence = sum(x[i] * relevance[i] for i in range(3))
pairs = sum(x[i] * x[j] * c
for (i, j), c in redundancy.items())
return -alpha * influence + 2 * (1 - alpha) * pairs
solutions = [(energy(x), x) for x in product((0, 1), repeat=3)]
print("Точный минимум учебной задачи:", min(solutions))
pi = 289006 / 694676
p, recall = 0.87, 0.97
acc = 1 - pi * (1 - recall) - pi * recall * (1 / p - 1)
print("Условная accuracy для бинарной модели:", acc)
sensitivity, fpr, prevalence = 0.97, 0.01, 0.001
ppv = sensitivity * prevalence / (
sensitivity * prevalence + fpr * (1 - prevalence))
print("PPV искусственного примера:", ppv)
16. Sources and limits of their verification
In-text citations lead to this provenance map. The numbers [14], [16], [25], [26] when discussing PDF bibliography refer to its internal list; they do not match the web/file identifiers.
- [file:1] Sie M.-F., Chang Y.-J., Lin C.-L., Chang C.-R., Liao S.-W. Efficient Bitcoin Address Classification Using Quantum-Inspired Feature Selection. Provided PDF, 21 pages, arXiv:2411.15425v1, 2024-11-23. Primary source: §2 — model; §3 — data and methods; table 3 — selection; tables 6–7 — metrics; §5 — feature origin. Preprint card .
- [file:2] “Quantum Technology Research Questions,” HTML provided, document date 10/30/2026. Methodological constructor, not an official standard; all 12 thematic sections, unification matrix, and checklist have been studied.
- [web:18] The arXiv card for the main preprint is arXiv:2411.15425 . The authors, date, abstract, and version history have been verified; the peer-reviewed status of this journal publication is not indicated by this card.
- [web:5] Lin Y.-J., Wu P.-W., Hsu C.-H., Tu I.-P., Liao S.-W. An Evaluation of Bitcoin Address Classification based on Transaction History Summarization. 2019, ICBC, pp. 302–310; arXiv:1903.07994 . Annotation directly tested: moments, eight classifiers, Micro-F1/Macro-F1 87%/86%.
- [web:15] Milne A., Rounds M., Goddard P. Optimal Feature Selection in Credit Scoring and Classification Using a Quantum Annealer. 1QBit white paper, 2017. Original source page . Verified by QUBO, German Credit Data, comparison with RFE, and indication of the 2018 extended version. Not confounded with independent hardware acceleration proof.
- [web:6] Hauke P., Katzgraber H.G., Lechner W., Nishimori H., Oliver W.D. Perspectives of quantum annealing: Methods and implementations. arXiv:1903.06559 , 2019. The domain description, authors, and associated DOI have been verified.
- [web:3] Journal version of Hauke et al.: Reports on Progress in Physics 83, 054401 (2020), DOI: 10.1088/1361-6633/ab85b8 . The preprint year 2019 should not be given as the journal publication year 2020.
- [web:34] Siemingfong/Quantom_Annealing . The available README was verified, but the full reproducibility of the source code and dataset was not. The repository text contains its own time estimates, which are not identical to the PDF experiment times.
Related publications known from the bibliography of the main PDF
- Toyoda K., Ohtsuki T., Mathiopoulos P.T. Multi-class bitcoin-enabled service identification based on transaction history summarization. 2018 IEEE iThings/GreenCom/CPSCom/SmartData, pp. 1153–1160. Reference [25] PDF; connection to underlying features confirmed by PDF conclusion, full text not independently verified. [file:1]
- Kirkpatrick S., Gelatt CD Jr., Vecchi MP. Optimization by simulated annealing. Science 220(4598), 671–680 (1983). Reference [16] PDF; used as a historical support for classical SA. [file:1]
