Quantum Support Vector Machines Kernel Generation with Classical Post Processing

We study the optimization of quantum kernel generation for support vector machine algorithms applied to data classification tasks. To improve classification efficiency, we implement classical post-processing techniques. The workflow begins with dimensionality reduction via Principal Component Analysis (PCA), preserving key features of high-dimensional datasets. We then construct the training kernel using the ZZ feature map. In the post-processing phase, overlaps with all quantum states are used, rather than only with the all-zero state as in standard quantum kernel methods. This allows kernel entries to be calculated as weighted sums of these overlaps, reducing the number of measurement shots needed and improving overall efficiency. We benchmark our method on the MNIST dataset for two classification problems: distinguishing digits ‘0’ vs. ‘1’ and separating even vs. odd digits. Comparing kernel scores—defined as the classification accuracy on unseen data—we find that our post-processed kernel consistently outperforms the standard quantum kernel, with the performance gap widening as the number of qubits increases.