Quantum-Machine-Learning-QML-Classifiers

Overview

This project implements and benchmarks a Hybrid Quantum–Classical Machine Learning (QML) classifier for a credit card fraud detection problem. The goal is to study the performance of variational quantum models on noisy, highly imbalanced, non-convex classification tasks and compare them against strong classical baselines under NISQ-era constraints.

The project emphasizes:

• Qubit-efficient data encoding
• Variational Quantum Circuits (VQCs)
• Hybrid quantum–classical training
• Fair benchmarking using standard ML metrics

---

Objectives

• Reduce high-dimensional classical data to a qubit-feasible representation
• Design a variational quantum classifier using Qiskit
• Train the model using hybrid optimization (SPSA)
• Benchmark quantum performance against classical ML models
• Analyze robustness and variability of QML under realistic constraints

---

Dataset

• Name: Credit Card Fraud Detection Dataset
• Samples: 100,000 transactions
• Features: 8 numerical features
• Target: Binary fraud label (highly imbalanced)

---

Methodology

Data Preprocessing

• Feature scaling using StandardScaler
• Missing value handling via median imputation
• Dimensionality reduction using PCA
• Feature reduction from 8 → 4 components to enable 4-qubit quantum encoding

Classical Baseline Models

• Support Vector Machine (RBF kernel)
• Random Forest Classifier
• Multi-Layer Perceptron (MLP)

Quantum Model

• Feature Map: ZZ Feature Map (4 qubits)
• Ansatz: Real Amplitudes variational circuit
• Quantum Neural Network: SamplerQNN
• Optimizer: SPSA (gradient-free, NISQ-friendly)

A balanced subsample of the training data is used for quantum training due to hardware and simulation constraints.

---

Training Strategy

• Hybrid quantum–classical optimization loop
• Classical optimizer updates quantum circuit parameters
• Quantum measurements used to generate predictions
• Performance evaluated on the full test set

Due to stochastic optimization (SPSA), results vary across runs. A representative run is reported.

---

Evaluation Metrics

• Accuracy
• AUC-ROC

---

Results Summary

Model	Accuracy	AUC-ROC
SVM (RBF)	~0.99	~0.998
Random Forest	~0.998	~0.999
MLP Neural Network	~0.996	~0.999
Hybrid Quantum ML	~0.05	~0.50–0.52

Classical models significantly outperform the quantum model on this dataset.
However, the quantum classifier demonstrates non-random discrimination capability, highlighting both the potential and current limitations of QML.

---

Reproducibility Notes

• SPSA optimization is stochastic
• Exact results may vary across runs
• A fixed random seed is used where possible
• ROC curves are generated from saved prediction outputs to ensure consistency

---

How to Run

Install Dependencies

pip install qiskit qiskit-machine-learning scikit-learn pandas numpy matplotlib

Execute

1. Open the Jupyter notebook
2. Run all the cells
3. Save prediction outputs immediately after training

⚠️Avoid rerunning training cells multiple times.