🛡️ Online Transaction Fraud Detection

📌 Project Overview

This project identifies fraudulent activity in financial transactions using a dataset of 6.3M+ records. The core challenge is the extreme class imbalance (only 0.13% are fraud). I implemented a multi-iteration approach using SMOTE and XGBoost to reach near-optimal detection levels.

📊 Key Results

Through 5 iterations of data engineering and balancing, the models achieved the following performance:

Iteration	Strategy	XGBoost AUC	Random Forest AUC
1	Baseline (Default Params)	87%	86%
2	Hyperparameter Tuning	99.5%	84.3%
3	SMOTE + Tuning	99.4%	98.7%
4	Undersampling + Tuning	98.8%	99.6%
5	SMOTE + Subsampling	99.0%	92.1%

🛠️ Feature Importance

The analysis revealed that the most influential features for detecting fraud are:

1. oldbalanceOrg (Balance before transaction)
2. newbalanceDest (Recipient's new balance)
3. amount (Size of the transaction)

📂 Project Structure

├── Fraud_detection.ipynb  # Main analysis and model training
├── requirements.txt       # Dependencies
├── .gitignore             # Excludes large data files
└── README.md              # Project documentation

🚀 How to Run

1. Clone the repo:

   git clone https://github.com/Akshat8510/Fraud-Detection-Project.git

2. Install dependencies:

   pip install -r requirements.txt

3. Run the Jupyter Notebook Fraud_detection.ipynb.