SVM vs. QSVM.

Overview

The context behind the code.

This experiment compares classical support vector machines with quantum-kernel classifiers on the Wisconsin Breast Cancer dataset. It explores what changes when the same reduced feature space is represented through two- and four-qubit quantum feature maps on an exact local simulator.

The problem

The problem

A quantum model's score is not meaningful without a carefully matched classical baseline and a transparent evaluation protocol. The experiment therefore examines preprocessing, kernel choice, tuning, runtime, and uncertainty together rather than treating one favorable result as evidence of an advantage.

Architecture

Architecture

Training-fitted scaling and PCA feed classical SVMs and ZZ-feature-map quantum kernels, followed by matched outer-test evaluation. The final phase uses five fixed stratified outer splits with five inner folds, saving candidate scores, model selections, predictions, runtime measurements, and paired analyses.

Submitted to Machine Learning: Science and Technology (IOP Publishing); not accepted. A simulator study, not a hardware benchmark or clinical application. Five overlapping splits and earlier feature-map selection limit inference; the report documents those limitations explicitly.

My contribution

My contribution

My exploration is recorded in an executable notebook, a supporting evaluation module, and validation tests. Saved candidate scores, model evaluations, and predictions make the experiment inspectable.

Engineering challenges

Engineering challenges

Preprocessing leakage, prior feature-map selection, and overlapping test splits all affect what the results can support. Exact statevector simulation also has a different cost model from quantum hardware, so kernel construction, model fitting, prediction, and tuning costs are recorded separately.

Engineering reflections

Engineering reflections

The reported classical baselines achieved higher mean F1 in this configuration, and the experiment does not demonstrate quantum advantage. The report recommends independent validation with a frozen protocol: the value is in understanding the evidence and its limits as carefully as the quantum model itself.

Technology

The tools behind the work.

• Python
• Qiskit
• scikit-learn
• NumPy
• SciPy
• Pandas
• Jupyter