A comparative study of Classical SVM vs Quantum SVM (QSVM) for detecting fraudulent credit card transactions, implemented using Qiskit and scikit-learn.

This project explores the potential of Quantum Machine Learning (QML) in a real-world binary classification task — credit card fraud detection. A ZZFeatureMap-based Quantum Kernel is used to train a QSVM on a PCA-reduced feature space, and its performance is compared against a tuned classical RBF-SVM.

PCA variance retained: 61.56% across 4 components
Test set size: 80 samples

The classical SVM significantly outperforms the QSVM in this experiment. The QSVM's lower performance is expected at this scale — quantum advantage typically emerges with higher-dimensional, more complex feature spaces that are difficult for classical kernels to separate.

The quantum feature map used is a ZZFeatureMap with reps=2 and linear entanglement across 4 qubits (one per PCA component). It encodes classical data into quantum states via parameterized rotation and entangling gates, enabling the quantum kernel to capture non-linear relationships in the feature space.

• Python 3.13
• qiskit — Quantum circuit construction
• qiskit-machine-learning — FidelityQuantumKernel
• qiskit-aer — Local quantum simulator
• scikit-learn — SVM, PCA, GridSearchCV, metrics
• pandas, numpy — Data handling
• matplotlib, seaborn — Visualization

qsvm-fraud-detection/
├── asset/
│   ├── classicalvsquantum_confusionmatrix.png
│   └── Quantum Kernel.png
├── data/
│   └── creditcard.csv          # Not included (see below)
├── qsvm_fraud.ipynb            # Main notebook
├── .gitignore
└── README.md

git clone https://github.com/your-username/qsvm-fraud-detection.git
cd qsvm-fraud-detection

python -m venv venv
source venv/bin/activate        # On Windows: venv\Scripts\activate

pip install qiskit qiskit-machine-learning qiskit-aer scikit-learn pandas numpy matplotlib seaborn jupyter

Download the Credit Card Fraud Detection dataset from Kaggle and place creditcard.csv inside the data/ folder.

jupyter notebook qsvm_fraud.ipynb

1. Data Loading & Balancing — Undersample the majority class (valid transactions) to match the minority class (fraud), resulting in a balanced 50/50 split.
2. Subset Sampling — Use 400 samples to keep quantum simulation tractable.
3. Train/Test Split — Stratified 80/20 split before any preprocessing (preventing data leakage).
4. Standardization — StandardScaler fit on training data only, then applied to test data.
5. PCA — Dimensionality reduction to 4 components, fit on training data only.
6. Classical SVM — Tuned RBF-SVM via GridSearchCV (C, gamma) with 5-fold cross-validation.
7. Quantum SVM — ZZFeatureMap (reps=2, linear entanglement) + FidelityQuantumKernel via Qiskit Aer simulator.
8. Evaluation — Accuracy, F1 Score, and confusion matrices for both models.

• Quantum simulation is slow — The QSVM uses a classical Aer simulator, not real quantum hardware. Kernel matrix computation scales as O(n²) in the number of training samples.
• Low PCA variance retained — Only ~61.56% of variance is preserved with 4 components, which impacts QSVM accuracy.
• Small dataset — 400 samples is a necessary trade-off for simulation feasibility.

This project is open-source and available under the MIT License.

Built with curiosity about the intersection of quantum computing and machine learning.
Feel free to open an issue or submit a PR!