Comparative Analysis of a Quantum SVM With an Optimized Kernel Versus Classical SVMs
Matheus Cammarosano Hidalgo
Abstract
SupportVector Machine (SVM) is a widely used algorithm for classification, valued for its flexibility with kernels that effectively handle non-linear problems and high-dimensional data. Businesses across industries face challenges in improving customer retention and reducing churn, making predictive models essential for identifying at-risk customers and enhancing revenue. This paper investigates an optimized quantum embedding kernel for SVM (Quantum SVM - QSVM) applied to a public bank customer dataset, featuring variables on customer relationships and churn indicators. While focused on the financial sector, the methodology is broadly applicable for reducing churn and boosting revenue across industries. QSVM performance is compared to SVMs with established kernels, including Radial Basis Function (RBF), linear, polynomial, and sigmoid. Experiments varied the number of variables from two to seven to evaluate their impact on model performance and kernel behavior. The experiments were conducted on quantum simulators, which faced scalability challenges addressed using reduced datasets. Even so, this study sheds light on the potential of QSVMs to effectively manage increasing numbers of variables in predictive models, offering valuable insights into their capability to handle complex, high-dimensional data and their applicability in real-world scenarios.