README_space.md

💳 Credit Card Fraud Detection — HF Space (Calibrated RF Model)

This is an interactive Gradio demo of a calibrated Random Forest model for credit card fraud detection.
The model was trained on the Kaggle Credit Card Fraud dataset,
and probability calibration ensures reliable decision thresholds for business scenarios.

---

🚀 How to Use

1. Upload your CSV with transaction rows.

   • Required columns: V1 … V28, Amount
   • Either include engineered features, or just add Time (seconds from start)
     → the app will automatically derive:
     • _log_amount
     • Hour_from_start_mod24
     • is_night_proxy
     • is_business_hours_proxy

2. Adjust the decision threshold with the slider.

   • Default is set to the validation threshold for Precision ≥90% (≈0.712).
   • Move it left/right to trade off between precision and recall.

3. Preview results (first 50 rows) or enable Return all rows for the full file.

   • Each row includes:
     • Fraud_Probability
     • Prediction (0 = normal, 1 = fraud)

4. Download results as predictions.csv.

---

🧪 Try with Example Data

You don’t need to bring your own data to test the app!
Just click Use Example inside the app, and it will load the included example_transactions.csv.

This file mimics the required structure:

• 60 transactions
• Columns: V1..V28, Amount, Time
• Probabilities + predictions are generated live with the same calibrated RF model.

---

📊 Notes

• The model is calibrated with Isotonic Regression for probability reliability.
• Default threshold corresponds to Precision ≥90%, aligning with fraud detection team priorities.
• For production use, re-tune thresholds regularly as data drift changes prevalence and costs.

---

🔗 Related

• Model repo on Hugging Face Hub
• Original Kaggle dataset