Financial Fraud Detection GNN

This project implements a Graph Neural Network (GNN) for detecting financial fraud using transaction data.

Overview

The project uses PyTorch and PyTorch Geometric to build and train a GNN model on transaction data. The goal is to classify transactions as fraudulent or non-fraudulent.

Features

• Graph-based representation of transaction data
• GNN model for classification
• Evaluation metrics including accuracy, precision, recall, and F1-score
• Visualization of training metrics and ROC curve

Installation

1. Clone the repository:

   git clone https://github.com/mmannan17/Financial_Fraud_DetectionGNN.git
   cd Financial_Fraud_DetectionGNN

2. Create and activate a virtual environment:

   python -m venv venv
   venv\Scripts\activate  # On Windows
   # or
   source venv/bin/activate  # On macOS/Linux

3. Install the required packages:

   pip install -r requirements.txt

Usage

1. Prepare your data files:

   • elliptic_txs_features.csv: Node features
   • elliptic_txs_edgelist.csv: Edge list
   • elliptic_txs_classes.csv: Class labels

2. Run the training script:

   python ff_gnn.py

3. View the results:

   • Confusion matrix: confusion_matrix.png
   • ROC curve: roc_curve.png
   • Training metrics: training_metrics.png

Results

During training, the model achieved the following performance metrics:

• Test Accuracy: 97.47%
• Test Precision: 97.47%
• Test Recall: 100.00%
• Test F1 Score: 98.72%

Confusion Matrix

Predicted Negative	Predicted Positive
0	169
0	6518

Training Progress

• Epoch 01: Loss: 2003620.6250, Val Loss: 668804.1875, Val Acc: 90.84%
• Epoch 02: Loss: 2680480.7500, Val Loss: 859666.5625, Val Acc: 90.84%
• Epoch 03: Loss: 3469554.5000, Val Loss: 836082.1250, Val Acc: 90.84%
• Epoch 04: Loss: 3422597.0000, Val Loss: 709417.6875, Val Acc: 90.84%
• Epoch 05: Loss: 2920227.5000, Val Loss: 541518.4375, Val Acc: 90.84%
• Epoch 06: Loss: 2217199.7500, Val Loss: 346993.0000, Val Acc: 90.84%

Early stopping was triggered during training.

Graph Data

• Edge Index Shape: [2, 234355]
• Max Index in Edge Index: 203768
• Min Index in Edge Index: 0
• Number of Nodes: 203769

Contributing

Contributions are welcome! Please open an issue or submit a pull request for any improvements or bug fixes.

License

This project is licensed under the MIT License. See the LICENSE file for details.

Contact

For questions or feedback, please contact Mustafa at [[email protected]].