Reproducible Benchmarks Reveal Instability of Quantum Kernel Methods versus Stability of Hybrid Variational Classifiers on NISQ Simulators

Author: Ahmad Rasidi¹¹ Independent Researcher; Undergraduate Student (3rd Semester, Bachelor’s Program), Bumigora University, Indonesia Abstract: Quantum Machine Learning (QML) stands at the frontier of near-term quantum computing applications, promising enhanced pattern recognition through quantum mechanical principles. However, empirical validation of QML models under Noisy Intermediate-Scale Quantum (NISQ) constraints remains fragmented and often lacks reproducibility and cross-platform consistency. This work presents a comprehensive benchmark comparing two canonical QML paradigms Hybrid Variational Quantum Classifier (VQC) and Quantum Kernel Support Vector Machine (QSVM) using a unified open-source framework built on PennyLane, PyTorch, Qiskit, and scikit-learn. The study introduces a reproducible, modular, and cross-device pipeline featuring YAML-driven configuration, deterministic seeding, automatic artifact logging, and support for CPU, CUDA, and MPS execution. Across six independent experimental runs (seeds 1001–1006), we evaluate test accuracy, F1-score, and ROC-AUC on a binary classification task. Results demonstrate that the Hybrid VQC achieves a mean test accuracy of 81.11% ± 5.75%, with consistently high F1-score (0.804 ± 0.055) and ROC-AUC (0.885 ± 0.060), showcasing robust performance despite optimization challenges inherent to variational algorithms. In stark contrast, the QSVM exhibits extreme instability: while achieving 50.3% accuracy in all runs (equivalent to random guessing on a balanced dataset), its F1-score varies dramatically, with a mean of 0.222 ± 0.365, primarily due to catastrophic failures (F1 = 0.000) in 4 out of 6 runs. This reveals the fragility of fixed quantum kernel methods under data variation, even for simple non-linear problems. The accompanying repository (GitHub – rasidi3112/Quantum-Machine-Learning) enables full reproducibility of all results and artifacts. This work establishes a transparent, statistically grounded benchmark for QML under NISQ constraints and highlights the critical trade-offs between the flexibility of hybrid approaches and the instability of kernel-based methods. “This document was partially refined using ChatGPT (OpenAI), and all content has been reviewed by the author.”