Credit Card Fraud Detection
Domain: Finance & Cybersecurity | Tools: Python, XGBoost, Scikit-learn, Matplotlib, Seaborn
Project Overview
This machine learning project investigates the detection of fraudulent credit card transactions using a real-world-style dataset. The goal was to develop and compare classification models and provide insights into transaction patterns and fraud detection techniques.
Objectives
- Train ML models to classify transactions as fraudulent or legitimate
- Conduct statistical analysis of fraud patterns
- Visualize trends by transaction type, location, and merchant
- Deploy a trained model for future use
Dataset Summary
- 98,000+ records with ~1% fraudulent transactions
- Features include transaction type, location, amount, merchant ID, and timestamps
- Date and time features engineered for modeling
Modeling
Three models were compared:
- Logistic Regression: 64% Accuracy | ROC-AUC: 0.66
- Random Forest: 97% Accuracy | ROC-AUC: 0.997
- Tuned XGBoost: 92% Accuracy | ROC-AUC: 0.976
| Model | Accuracy | ROC-AUC | F1 Score |
|---|---|---|---|
| Logistic Regression | 64% | 0.66 | 0.64 |
| Random Forest | 97% | 0.997 | 0.97 |
| Tuned XGBoost | 92% | 0.976 | 0.92 |
Feature Importance
Top predictors included transaction type, amount, merchant ID, and time-of-day features.
Deployment
The Random Forest and XGBoost models were serialized using Joblib for deployment in application as a Flask API. Future enhancements include real-time fraud scoring and user feedback loops.