Credit Card Fraud Detection

This repository implements a Credit Card Fraud Detection pipeline using XGBoost, SHAP for model interpretation, and Streamlit for interactive demonstration. It is organized into:

• notebooks/: Jupyter notebooks covering end-to-end data exploration, feature engineering, model training, tuning, and interpretation.

• scripts/: Python utility modules:

  • process_final.py: polprocess(df, cat_rates) for consistent feature engineering.

• artifacts/: Serialized model & metadata for deployment:

  • fraud_slim.json: trained slim XGBoost Booster.
  • slim_features.joblib: list of features expected by the slim model.
  • category_rates.joblib: mapping of category → historical fraud rate.
  • le_category.joblib: LabelEncoder for transaction categories.
  • uf_names.joblib: user‑friendly category names.

• streamlit_app/: a polished Streamlit application (pol_app.py) that loads artifacts and lets you simulate new transactions.

---

🔧 Installation

1. Clone the repo

   git clone https://github.com/<your‑username>/CC_Fraud_Detection.git
   cd CC_Fraud_Detection

2. Create & activate a Python environment (conda or venv):

   conda create -n fraud-detect python=3.11
   conda activate fraud-detect

3. Install dependencies

   pip install -r requirements.txt

📓 Jupyter Notebook

• Open notebooks/CC_Fraud_polished.ipynb to step through:

  1. Setup & Data Load: imports, Kaggle download, preview.
  2. Preprocessing & Feature Engineering: apply polprocess() to train/test splits.
  3. Hyperparameter Tuning & Training: train slim XGBoost with early stopping.
  4. Model Interpretation: SHAP summary & beeswarm plots.
  5. Export Artifacts: save model, feature list, encoders to artifacts/.

Run all cells to reproduce results and regenerate the artifacts/ files.

🚀 Streamlit App

1. Launch:

   streamlit run streamlit_app/pol_app.py

2. Simulate new transactions by adjusting amount, category, hour, and population.

3. Threshold slider dynamically changes fraud vs. legitimate decision cutoff.

🗂️ Folder Structure

├── artifacts/               # trained models + metadata
│   ├── fraud_slim.json
│   ├── slim_features.joblib
│   └── ...
├── notebooks/
│   ├── CC_Fraud_polished.ipynb
│   └── NB_archive/...
├── scripts/
│   ├── process_final.py
│   └── booster_wrapper.py
├── streamlit_app/
│   └── pol_app.py
├── requirements.txt
└── README.md

📈 Results Summary

• Slim Model: 12‑feature XGBoost, AUC ≈ 0.99, PR‑AUC ≈ 0.80, F1 optimized at threshold ≈ 0.70.
• Interpretation: SHAP identifies top drivers (amount, category TE, hour patterns, etc.).

📂 Deployment

All required artifacts are in artifacts/. To deploy:

1. Ensure artifacts/ is alongside streamlit_app/.
2. Run the app as above.

⚖️ License

This project is MIT‑licensed—feel free to reuse and adapt in your portfolio or demos.