Fraud Detection System

Overview

This project is a fraud detection system an internal tool that combines a LightGBM (Gradient Boosting Decision Trees) model with a Neural Network for enhanced performance. The system uses a hybrid approach to predict fraudulent transactions, and the models are containerized using Docker for easy deployment. Additionally, LIME is used for explainability, helping to understand which features are most important in predicting fraud.

Features

• Hybrid Model: Combines a LightGBM model with a Neural Network.
• LIME Explainability: Interpret individual predictions.
• Model Calibration: LightGBM is calibrated to improve probability outputs.
• Precision-Recall Optimization: Custom threshold optimization for improved precision-recall tradeoff.
• Dockerized: Containerization for easy replication and future deployment.

Project Setup

To set up the project locally, follow these steps:

1. Clone the Repository:

   git clone https://github.com/Issac-Kondreddy/fraud-detection-system.git
   cd fraud-detection-system

2. Install Dependencies:

   • Install Python packages:

     pip install -r requirements.txt

3. Run Django Migrations:

   python manage.py migrate

4. Run the Django Server:

   python manage.py runserver

---

Usage

After the server is running, you can make predictions by sending a POST request to the /predict/ endpoint.

Example Request:

curl -X POST http://127.0.0.1:8000/predict/ \
-H 'Content-Type: application/json' \
-d '{
  "feature1": value1,
  "feature2": value2,
  ...
}'
Example Response:
```bash
{
  "prediction": "Fraud" or "Not Fraud",
  "probability": 0.85
}

Model Information

LightGBM Model: This model uses 400 estimators, a max depth of 15, and balanced class weights to handle the imbalanced nature of fraud detection data. It is calibrated using sigmoid calibration. Neural Network: A simple 3-layer neural network with Dropout and ReLU activation, trained on a hybrid dataset (original features + LightGBM output). LIME Explainability: LIME provides feature importance on a per-instance basis for enhanced transparency in model decisions. Model Performance The hybrid model (LightGBM + Neural Network) has the following performance metrics on the test data:

Accuracy: 1.000 Precision: 0.999 Recall: 1.000 F1-Score: 1.000 Optimal Threshold: 0.85 (calculated from precision-recall curve).

Confusion Matrix:

	Predicted Not Fraud	Predicted Fraud
Actual Not Fraud	62675	22
Actual Fraud	4	34123

---

ROC Curve:

The model achieves a high AUC-ROC score, demonstrating its ability to discriminate between fraudulent and non-fraudulent transactions.

---

Precision-Recall Curve:

Optimal precision-recall tradeoff was identified using a threshold of 0.85.

---

API Endpoints

• /predict/: POST request to predict whether a transaction is fraudulent.
  • Input: JSON with the required features.
  • Output: JSON with the prediction ("Fraud" or "Not Fraud") and the confidence score.

---

Docker Instructions

To run the project using Docker:

1. Build the Docker image:

   docker build -t fraud-detection-system .

2. Run the Docker container:

   docker run -p 8000:8000 fraud-detection-system

The API will now be accessible at http://127.0.0.1:8000.

If using docker-compose, simply run:

docker-compose up

Technologies Used

Python 3.x LightGBM: Gradient boosting model for fraud detection. Neural Network (Keras): A fully connected neural network for hybrid model prediction. LIME: Used for explainability of individual predictions. Django: Backend web framework to serve the model. Docker: For containerizing the project.

Future Improvements

1. Deploy to Cloud: Future deployment to AWS, GCP, or Azure.
2. Real-Time Predictions: Add streaming support for real-time fraud detection.
3. Advanced Tuning: Further hyperparameter tuning or exploring advanced models (e.g., XGBoost).