Credit Card Fraud Detection System

End-to-end Machine Learning pipeline for detecting fraudulent credit card transactions
Built with focus on imbalanced data handling, production-grade practices, and deployable API.

API DEMO

Project Highlights

• Complete ML pipeline: from raw data to live prediction API
• Effective handling of extreme class imbalance (~0.17% fraud cases)
• Threshold tuning for optimal precision-recall trade-off in production
• Stratified cross-validation + model comparison
• FastAPI inference service
• Explainability support feature importance

Best model: XGBoost
Achieved excellent fraud-class performance after threshold optimization.

Dataset

Credit Card Fraud Detection (Anonymized transactions – September 2013)
→ Source: Kaggle – Credit Card Fraud Detection

Features:

• Time – seconds since first transaction
• V1–V28 – PCA-transformed features
• Amount – transaction amount
• Class – target (0 = normal, 1 = fraud)

Key statistics:

• Total transactions: 284,807
• Fraud cases: 492 (~0.172% – highly imbalanced)

Download the dataset

1. Go to: https://www.kaggle.com/datasets/mlg-ulb/creditcardfraud
2. Sign in / create a free Kaggle account
3. Click Download (file: creditcard.csv ~150 MB)
4. Place the file in the data/ folder (create if needed)

   Note: The dataset is not included in this repository due to size and licensing.

Project Structure

fraud_detection/
├── data/                    # Put creditcard.csv here
├── notebooks/
│   └── 01_eda.ipynb         # EDA
├── src/
│   ├── preprocessing.py
│   ├── models.py
│   ├── train.py
│   ├── predict.py
│   ├── evaluation.py
│   ├── interpret.py
│   └── api.py
├── models/
│   ├── best_model.pkl
│   ├── best_scaler.pkl
│   └── best_threshold.pkl
├── requirements.txt
├── .gitignore
└── README.md

Installation

1. Installation

# Clone repository
git clone https://github.com/YOUR-USERNAME/credit-card-fraud-detection.git
cd credit-card-fraud-detection

# Create & activate virtual environment
python -m venv venv
source venv/bin/activate    # Linux / macOS
# or
venv\Scripts\activate       # Windows

# Install dependencies
pip install -r requirements.txt

2. Train the model

cd src
python train.py

3. Run API

uvicorn src.api:app --reload --port 8000

→ Open: http://127.0.0.1:8000/docs

Model Performance (Best – XGBoost)

Metric	Value
Precision (Fraud)	0.96
Recall (Fraud)	0.78
PR-AUC	0.84

License

MIT License – feel free to use this project for learning purposes.