A proposed quantum kernel ranked first on mean balanced accuracy in every evaluated configuration across two synthetic settings, according to a new preprint. The result held against conventional kernels and the InterFea baseline, although the margin over InterFea was close in the smaller setting and modest but consistent in the larger one. The paper's hypothesis is that when predictive structure is defined by interactions among features, a similarity measure built to reflect those interactions may represent the task better than distance alone. The result is conditional: on two fraud benchmarks, the quantum method led one and ranked second on the other.
Building similarity around interactions
The method is designed for cases where distance between observations is not enough to capture predictive structure. Its feature map keeps coordinate information with one-body phases and adds a block-wise Pauli-string phase for each block's high-order product. The paper states that the resulting quantum kernel is positive semidefinite. The study compared it with conventional distance-based kernels and the InterFea baseline. Kernel methods used precomputed Gram matrices with kernel ridge regression; hyperparameters were selected on validation data, and the selected model was then refitted on the full training set.
Each synthetic run used 2,000 training observations, split into 1,600 subtraining observations and 400 validation observations, plus an independent 500-observation test set. Configurations were repeated over 10 random seeds to compare mean performance.
The advantage was clearest in designed tests
The method's edge persisted in the study's higher-order scaling runs. At 60 dimensions with sixth-order interactions, it recorded mean accuracy of 0.706 and mean F1 of 0.412. At 80 dimensions with eighth-order interactions, the figures were 0.695 and 0.395. It ranked first on both measures in both settings, with mean-F1 gaps over InterFea of 0.032 and 0.075, respectively.
Those paired comparisons were reported with 95% confidence intervals for the quantum-minus-InterFea F1 difference of 0.011 to 0.053 in the 60-dimensional case and 0.032 to 0.119 in the 80-dimensional case. The corresponding two-sided p-values were 0.007 and 0.004. The authors do not frame these findings as evidence of universal dominance.
The advantage was also visible as the synthetic training set changed size. The quantum kernel had the highest mean test accuracy at every investigated size, rising from 0.718 with N = 250 to 0.753 with N = 2,000. InterFea rose from 0.666 to 0.734 over the same range.
Fraud benchmarks gave a mixed verdict
On the Credit Card Fraud Detection benchmark, the quantum kernel ranked first on both mean accuracy and mean F1. Its mean F1 was 0.472, compared with 0.348 for the Poly2 kernel - an absolute gap of 0.124, or about 36% relative. The benchmark used a fixed chronological test set of 56,962 observations; the original training partition had 227,845 observations, with 834 resampled for model fitting.
The picture changed on numeric-only IEEE-CIS Fraud Detection. Laplacian ranked first, with mean accuracy of 0.831 and F1 of 0.267; the quantum kernel ranked second, at 0.821 accuracy and 0.239 F1. The experiment used 60 features, grouped as 10 blocks for sixth-order interactions, and resampled 6,000 training observations from a larger 561,013-observation training partition; its test partition contained 29,527 observations.
What the study does not claim
The contrast is important to how the authors interpret the work. They see the synthetic ranking as evidence of alignment between feature-map geometry and interaction structure, while the fraud results indicate conditional competitiveness rather than universal dominance. They describe the contribution as predictive and representational, not as a computational quantum advantage.
The block-factorized kernel has an exact classical evaluation, and the circuit resources reported by the paper describe a logical, pre-transpilation representation rather than hardware-executed quantum computation. The evaluation therefore does not establish computational quantum speedup.
The synthetic data are generated programmatically, while the real benchmark datasets are publicly available through their respective sources and the Fraud Dataset Benchmark. The authors say code is available on reasonable request. The manuscript is an arXiv preprint, identified in the supplied metadata as arXiv:2608.24631v1 and dated 25 August 2026; no journal publication is identified there.
Paper data and sources
Original title: When Similarity Is Interaction-Driven: Quantum Kernels for Regime-Sensitive Learning
Authors: Hanqiu Peng, Jianlong Lu, Ying Chen
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text