Empirical Evaluation of Quantum Support Vector Machines with
Shallow Feature Maps on Classical Benchmark Datasets

Abstract: Quantum machine learning seeks to integrate quantum-mechanical principles with statistical inference to construct models capable of representing complex data beyond the limits of classical feature mappings. This study investigates whether quantum support vector machines (QSVMs) provide measurable performance or representational advantages over established classical classifiers on standard biomedical benchmark datasets. Quantum and classical models were trained and tested under equivalent preprocessing and validation protocols to ensure methodological parity in the controlled evaluation. QSVMs are implemented using parameterized quantum feature maps and evaluated through kernel-based classification. A

support vector classifier (SVC) was trained with a precomputed kernel, with kernel entries computed using a statevector simulator to isolate representational capacity from sampling and hardware noise. Across datasets, optimized classical models consistently achieve higher

predictive accuracy than their quantum counterparts, although QSVMs perform competitively in lower-dimensional regimes. Qualitative analysis of the resulting kernel matrices reveals structured off-diagonal correlations indicative of higher-order feature interactions introduced by quantum entanglement, suggesting that QSVMs capture non-separable relationships that are not directly accessible to standard Gaussian kernels. Despite this richer representational structure, current simulated QSVM implementations do not surpass tuned classical baselines in overall classification accuracy. These findings indicate that the primary near-term value of QSVMs lies in their distinctive feature-space geometry rather than in immediate performance gains, and that QSVMs implicitly learn a richer set of feature interactions from circuit topology rather than through explicit hyperparameter optimization. This highlights the importance of improved encoding strategies and hardware-aware optimization for realizing practical quantum advantage in supervised learning.