The financial-transaction feature map FT-FM, introduced in the first paper of this series, is empirically more expressive than generic quantum encodings but operates in a near-maximally entangled regime in which quantum-kernel values concentrate and kernel-based classification fails. This paper introduces FT-FM-lite, a refinement of FT-FM with a single entanglement-strength hyperparameter β ∈ [0, 1] that preserves the domain-informed encoding axioms while allowing the feature map to operate in the intermediate expressibility–trainability regime; β = 1 recovers FT-FM exactly. A validation-based protocol selects β per training split. In a ten-seed synthetic diagnostic against exactly implemented matched controls, validation-selected FT-FM-lite attains a mean held-out F1 of 0.620 (95% bootstrap CI [0.530, 0.709]), exceeding exact amplitude encoding (0.362) and Qiskit’s ZZFeatureMap (0.033) with Holm–Bonferroni-corrected significance, while its advantage over full-strength FT-FM (0.490) is not significant after correction. We further specify a leakage-controlled, case-control benchmarking protocol with paired Wilcoxon significance testing and apply it to the Kaggle Credit-Card Fraud dataset over ten repeated stratified case-control splits. In contrast to the synthetic diagnostic, validation selection chose full strength (β = 1) in every seed, and FT-FM-lite attained a mean case-control test F1 of 0.891 (95% bootstrap CI [0.870, 0.907]; ROC-AUC 0.971; PR-AUC 0.976) — significantly above exact amplitude encoding and significantly below a matched RBF-SVM after Holm–Bonferroni correction, and statistically indistinguishable from Qiskit’s ZZFeatureMap and the remaining classical baselines. Balanced-cohort metrics are benchmark values, not deployment-prevalence estimates.
The financial-transaction feature map FT-FM, introduced in the first paper of this series, is empirically more expressive than generic quantum encodings but operates in a near-maximally entangled regime in which quantum-kernel values concentrate and kernel-based classification fails. This paper introduces FT-FM-lite, a refinement of FT-FM with a single entanglement-strength hyperparameter β ∈ [0, 1] that preserves the domain-informed encoding axioms while allowing the feature map to operate in the intermediate expressibility–trainability regime; β = 1 recovers FT-FM exactly. A validation-based protocol selects β per training split. In a ten-seed synthetic diagnostic against exactly implemented matched controls, validation-selected FT-FM-lite attains a mean held-out F1 of 0.620 (95% bootstrap CI [0.530, 0.709]), exceeding exact amplitude encoding (0.362) and Qiskit’s ZZFeatureMap (0.033) with Holm–Bonferroni-corrected significance, while its advantage over full-strength FT-FM (0.490) is not significant after correction. We further specify a leakage-controlled, case-control benchmarking protocol with paired Wilcoxon significance testing and apply it to the Kaggle Credit-Card Fraud dataset over ten repeated stratified case-control splits. In contrast to the synthetic diagnostic, validation selection chose full strength (β = 1) in every seed, and FT-FM-lite attained a mean case-control test F1 of 0.891 (95% bootstrap CI [0.870, 0.907]; ROC-AUC 0.971; PR-AUC 0.976) — significantly above exact amplitude encoding and significantly below a matched RBF-SVM after Holm–Bonferroni correction, and statistically indistinguishable from Qiskit’s ZZFeatureMap and the remaining classical baselines. Balanced-cohort metrics are benchmark values, not deployment-prevalence estimates.
Claims about quantum machine learning are routinely established on a single dataset and generalised informally. This paper tests one such claim directly. Earlier work in this series found, under a leakage-controlled predeclared protocol, that a financial-transaction-specific quantum feature map does not outperform matched classical baselines on the Kaggle Credit-Card Fraud dataset. We freeze source-column contracts and an experimental protocol in advance, inherit every predicted direction from that published result, and replicate on CICIDS-2017 (2.8 million flows) and UNSW-NB15 (257,673 records) — six paired comparisons under one Holm–Bonferroni correction. The replication is partial, and the failure is informative. The map’s advantage over naive amplitude encoding replicates on both datasets and is now established across three datasets and two domains. But the prediction that a matched RBF-SVM would win is significantly reversed on UNSW-NB15, where the quantum kernel wins by a median 0.021 F1 — identifying a regime in which the approach pays. That result is bounded: a random forest on identical cohorts outperforms the quantum kernel on both datasets, so no general claim of quantum advantage is supportable. Two further findings weaken the premise of domain-specific encoding itself. The advantage over generic quantum encodings differs in sign between the two network datasets, and the unanimous entanglement strength selected on financial data does not reproduce on either. We conclude that quantum kernel results reported on one dataset should not be assumed to transfer, in either direction, and that predeclared cross-domain replication is cheap enough to be routine. Balanced-cohort metrics are benchmark values, not deployment-prevalence estimates.
Preliminary empirical characterisation on synthetic financial-transaction data confirms FT-FM's predicted expressibility advantage over generic feature maps (Kullback– Leibler divergence to the Haar distribution of 4.01 ± 1.16 versus 21.24 ± 0.79 for a ZZ-like baseline and 19.53 ± 0.43 for an amplitude-like baseline at matched qubit count), and exposes a kernel-concentration phenomenon (Meyer–Wallach Q = 0.948 ± 0.004) that motivates a refined construction (FT-FM-lite) with a tunable entanglement-strength hyperparameter; full empirical benchmarking on Kaggle CCF, CICIDS-2017, and UNSW-NB15 is the subject of a forthcoming paper in this series.
Entanglement entropy is an attractive candidate for explaining quantum machine-learning models: it is intrinsic to the encoding, cheap to compute, and requires no measurement. We test whether it is a faithful feature attribution for a financial-transaction quantum feature map, under a protocol predeclared down to its thresholds, its comparison family, and its posture on failure. Two questions are separated by construction. First, does the operating point selected for accuracy leave usable attribution signal? Second, does that signal correspond to what actually drives the model’s decisions? On the Kaggle Credit-Card Fraud dataset over ten repeated case-control splits, per-feature entropy dispersion is 71% lower at the accuracy-selected entanglement strength than at the midpoint of the grid, and the difference is significant after Holm–Bonferroni correction; however, a lower-strength regime with 3.4× the dispersion is available at no statistically detectable accuracy cost. Faithfulness fails decisively. Entropy rankings are significantly anti-correlated with model-aware occlusion attributions (median Spearman −0.299), and on a synthetic task whose informative features are known by construction and which the quantum model demonstrably learns, occlusion recovers ground truth (+0.237) while entropy does not (−0.025), both significant after correction. A model-randomisation check identifies the cause exactly: per-feature entropy is invariant to the trained model, with maximum change of exactly zero across all ten seeds, so it cannot distinguish a correctly trained model from one fitted to shuffled labels. We conclude that entanglement entropy is a genuine and inexpensive diagnostic of the encoding but not an explanation of the model, and we report this refutation as predeclared. Balanced-cohort metrics are benchmark values, not deployment-prevalence estimates.
Financial fraud detection remains difficult due to extreme class imbalance and subtle, nonlinear patterns in transactional data. We investigate a hybrid quantum–classical approach that couples a fidelity-based quantum kernel with a logistic regression classifier (QEKLR). A custom ZZ-inspired feature map comprising Hadamards, mixed-axis Ry/Rzencodings, tunable Ising-ZZ cross terms, and data re-uploading, embeds PCA-reduced, Min–Max scaled features into quantum states. The resulting fidelity kernel is evaluated and used within kernel logistic regression. Experiments on two benchmarks: BankSim (synthetic) and the European Credit Card Fraud dataset, use balanced subsets created via Random Under-Sampling. The proposed model attains the best performance on BankSim (AUC = 0.96; strongest F1), and is competitive on the credit-card dataset (AUC = 0.94), closely tracking XGBoost (AUC = 0.96) while outperforming logistic regression and a variational quantum classifier baseline. These results indicate that quantum kernel methods can provide practical gains on tabular fraud tasks while remaining aligned with current NISQ constraints. We also discuss the quadratic cost of exact kernel evaluation and outline scalable approximations for future deployment on larger, highly imbalanced datasets.
Detecting credit card fraud is notoriously difficult because fraudulent charges make up less than 0.2% of total transactions. Quantum Machine Learning (QML), particularly Quantum Kernel Support Vector Machines (QSVM), promises richer feature representations by embedding classical data into high-dimensional Hilbert spaces. However, whether this theoretical advantage holds up when managing real-world financial data bottlenecks remains unclear. We designed a controlled benchmark comparing three quantum encodings, Angle, ZZFeatureMap, and IQP against classical models while keeping sample sizes, feature counts, and scaling strictly matched. To handle class skew without corrupting quantum state overlaps, we embedded synthetic minority oversampling (SMOTE) directly into the quantum Gram matrix workflow, raising PR-AUC scores by roughly +0.04 across feature maps. When trained on the full European credit card dataset with 30 features, classical baselines reached a top PR-AUC of 0.983. However, restricting classical models to match QSVM hardware constraints (500 samples, 4 PCA features) dropped their PR-AUC to 0.750. This puts our best quantum result 0.730 PR-AUC using IQP encoding on 150 samples into realistic context. Testing on physical quantum hardware reduced performance further to 0.62 due to CNOT gate errors and thermal noise. Overall, our findings pinpoint two main operational bottlenecks for near-term quantum kernels: the scaling of Gram matrix evaluation and heavy information loss from squeezing features onto small qubit registers.
Quantum machine learning on near-term noisy quantum devices has generated substantial theoretical interest, but rigorous empirical comparisons under realistic noise on practically relevant data remain scarce. This paper compares two paradigmatic quantum learning models, a Quantum Support Vector Machine (QSVM) built on the ZZFeatureMap quantum kernel and a Variational Quantum Classifier (VQC) with an EfficientSU2/RealAmplitudes ansatz, against tuned classical baselines (SVM with four kernels, Random Forest, XGBoost, LightGBM and CatBoost) on the ULB Credit Card Fraud dataset (284,807 transactions, 0.17% fraud). All models share an identical 4-qubit PCA-reduced feature space, evaluated on the full unbalanced test fold over 15 fits (3 folds × 5 seeds) and reported as mean ± standard deviation with bootstrap confidence intervals, AUPRC as the primary metric. Noise robustness is assessed under depolarizing noise p∈{0,0.001,0.01,0.05}, with ranking preservation measured directly through Spearman ρ and Kendall τ between the noisy and noiseless decision scores rather than read off AUPRC, alongside the per-paradigm computational cost. At four qubits the classical baselines lead (AUPRC 0.60 to 0.74, CatBoost best), above the VQC (0.494) and the QSVM (0.240); the controlled QSVM-versus-RBF–SVM comparison puts the cost of the quantum kernel at about 0.45 AUPRC. Under noise the QSVM keeps its score ranking (ρ=0.998 at p=0.001, 0.906 at p=0.01) and an operational decision threshold (recall 0.87 to 0.89, stable calibration), while the VQC AUPRC peaks non-monotonically at p=0.01 (0.494 rising to 0.654, then easing to 0.569 at p=0.05) even as its ranking decays monotonically (ρ from 0.72 to near zero), so average precision on its own misrepresents how noise affects it. The quantum models do not surpass the tuned classical reference at four qubits; the contribution is methodological: under noise, AUPRC has to be read together with a genuine rank statistic, because the two can move in opposite directions.