I Dockerized a Bank-Grade Credit Card Fraud Detection App with XGBoost (Recall 0.92 + SHAP)

#showdev #machinelearning #docker #xgboost

I Dockerized a Bank-Grade Credit Card Fraud Detection App with XGBoost

1. Project Background & Challenge

Credit card fraud detection is a classic extreme imbalanced data problem (fraud rate only 0.172%).

Normal accuracy looks amazing (~99.8%), but in real business the cost of false negatives is huge.

So I built the model with a Recall-first approach.

2. Key Results

Recall: 0.92
PR-AUC: 0.85
SHAP analysis clearly identified V14 and V17 as the top fraud drivers

3. Tech Stack & Production Features

Model: XGBoost + scale_pos_weight for imbalance
Production: Docker + docker-compose
Testing: Full unit tests
Model persistence: joblib with DataFrame input
Dependencies: All strictly pinned

4. How to Run (Docker)

docker compose up

5. GitHub Repository

https://github.com/Retro099/ML-Projects/tree/main/Credit_Card_Fraud_Detection

6. What I Learned

Always use Recall + PR-AUC as main metrics for imbalanced data
In production, always use DataFrame for predictions
Dockerization dramatically increases portfolio credibility

If you found this project useful, feel free to star the repository ⭐

DEV Community