Streaming ML pipeline detecting credit card fraud in real time XGBoost + Isolation Forest ensemble | AUC-ROC 0.9833 | Live dashboard
Credit card fraud costs billions annually, and most fraud detection systems operate in batch, catching fraud hours or days after it happens. This project builds a real-time streaming system that scores every transaction the instant it occurs, with a live operations dashboard for monitoring.
| Model | Metric | Score |
|---|---|---|
| XGBoost | AUC-ROC | 0.9833 |
| XGBoost | Avg Precision | 0.8819 |
| XGBoost | Recall (fraud) | 80% |
| XGBoost | Precision (fraud) | 96% |
| Isolation Forest | AUC-ROC | 0.9491 (unsupervised) |
Dataset: 284,807 transactions, 492 fraud (0.17% fraud rate) — extreme class imbalance handled with SMOTE oversampling.
- Fraud transactions average $122 vs $88 for legit — higher value, contrary to common assumption
- Top SHAP features: V14, V4, V12, V10, V17 (PCA-transformed signals)
- 154 of 492 frauds (31%) occur at night (10pm-5am)
The Streamlit dashboard shows: live transaction throughput, threat level indicator, scrolling alert ticker, fraud score distribution, and merchant category breakdown — all updating in real time from the Redis stream.
Transaction producer (50 tx/sec)
-> Redis queue (txn:queue)
-> Consumer worker
-> XGBoost + Isolation Forest ensemble scoring
-> SHAP explanation for flagged transactions
-> Redis alerts store
-> FastAPI (/predict, /metrics, /alerts)
-> Streamlit live dashboard
All 5 services containerized with Docker Compose — one command deploys the entire system.
Python | XGBoost | Isolation Forest | SHAP | SMOTE Redis | FastAPI | Streamlit | Plotly | Docker Compose | MLflow
git clone https://github.com/Efrrowini/fraud-detection
cd fraud-detection
docker-compose up --buildAPI docs: http://localhost:8000/docs Dashboard: http://localhost:8501
python -m venv venv
venv\Scripts\activate
pip install -r requirements.txt
python -m src.features
python -m src.train
python -m src.anomaly
# In separate terminals:
python streaming/producer.py
python streaming/consumer.py
streamlit run app/dashboard.pyKaggle Credit Card Fraud Detection — 284,807 European cardholder transactions, September 2013. Features are PCA-transformed for privacy (V1-V28) plus Time and Amount. Source
Built by Efrrowini | Presidency University Bangalore | Data Science Portfolio