Detecting Fraudulent Credit Card Transactions

This project is a machine learning-driven fraud detection system that analyzes customer transaction patterns to classify whether a credit card transaction is legitimate or fraudulent. Built using Python and core data science libraries, it simulates real-world financial fraud detection scenarios with over 1 million records and no missing values.

Objective

To develop and evaluate supervised machine learning models to accurately classify fraudulent credit card transactions and explore the impact of various transaction behaviors on the likelihood of fraud.

Dataset Overview

• Source: Kaggle
• Size: 1 million transaction records
• Target Variable: fraud (binary: 1 = fraudulent, 0 = legitimate)
• Key Features:
  • distance_from_home
  • distance_from_last_transaction
  • ratio_to_median_purchase_price
  • used_pin_number, used_chip, online_order

Research Hypotheses

1. Online transactions are more likely to be fraudulent.
2. Greater distance from home increases the chance of fraud.
3. Long gaps between transaction locations might indicate fraud.

Tech Stack

• Language: Python
• Libraries: Pandas, NumPy, Matplotlib, Scikit-learn
• Modeling: Logistic Regression, K-Nearest Neighbors (KNN), Random Forest, SVC

Project Structure

Fraud-Detection-ML/
│
├── data/
│   └── card_transdata.csv                  # Raw transaction dataset
│
├── notebooks/
│   ├── SmartQuestion-1.py                  # EDA and hypothesis testing
│   ├── SmartQuestion-3.py                  # Classification models
│   ├── SmartQuestion-4.py                  # Model improvement
│   └── Fraudulent_Credit_Detection.py      # Consolidated fraud detection code
│
├── src/
│   ├── get_data.py                         # Load and preprocess data
│   ├── split_data.py                       # Train-test split utility
│   ├── build_model.py                      # Model training module
│   ├── model_metrics.py                    # Evaluation metrics
│   └── master.py                           # Main controller script
│
└── README.md                               # You are here

Model Performance

Model	Accuracy	Recall	F1 Score
Logistic Regression	0.972	0.842	0.902
KNN (k=10)	0.997	0.978	0.983
Random Forest	1.000*	1.000	1.000
SVC	0.991	0.660	0.774

Perfect training score; may require tuning or further evaluation for real-world scenarios.

Key Takeaways

• KNN with K=10 was the most balanced and effective model.
• Random Forest hit perfect metrics but may be overfitting.
• Online orders and distance features played a significant role in fraud prediction.

Author

Surya Vamsi Patiballa
MS in Data Science – George Washington University

• Email :- svamsi2002@gmail.com
• LinkedIn :- https://www.linkedin.com/in/surya-patiballa-b724851aa/
• Resume :- https://drive.google.com/file/d/178IYcArC6YYVdJiIwRmJYodzKZ-JXe-D/view?usp=sharing

Future Enhancements

• Real-time prediction web app integration
• Auto-retraining pipelines
• Feature engineering from time-based patterns