Fraud Detection in Bitcoin Transactions using Graph Convolutional Networks (GCNs)

---

Project Description

bitcoin_fraud_detection is a project aimed at detecting fraudulent Bitcoin transactions using Graph Convolutional Networks (GCNs). The project leverages the Elliptic dataset and combines the strengths of C++ for data preprocessing and Python for implementing and training the GCN model. This hybrid approach ensures efficient data handling and powerful machine learning capabilities.

---

Features

• Data Preprocessing in C++: Efficient parsing and cleaning of transaction data.
• Graph Construction: Creation of a transaction graph using NetworkX.
• Graph Neural Network (GNN): Implementation of a GNN using PyTorch Geometric for fraud detection.
• Visualization: Visualization of transaction graphs and model performance metrics using Plotly.

---

Project Structure

bitcoin_fraud_detection/
│
├── data/
│   ├── filtered/
│   │   ├── filtered_classes.csv
│   │   ├── filtered_edgelist.csv
│   │   └── filtered_features.csv
│   └── unfiltered/
│       ├── elliptic_txs_classes.csv
│       ├── elliptic_txs_edgelist.csv
│       └── elliptic_txs_features.csv
│
├── src/
│   ├── data_preprocessing.cpp
│   └── CMakeLists.txt
│
├── training/
│   ├── data_preparation.ipynb
│   ├── gcn_model_weights.pth
│   ├── graph_data.pt
│   └── training.ipynb
│
├── visualization/
│   ├── data_plot.png
│   ├── data_predictions_plot.png
│   └── data_visualization.ipynb
│
├── README.md
└── LICENSE

Setup Instructions

C++ Environment Setup

1. Compile the C++ Code:

   cd src
   mkdir build
   cd build
   cmake ..
   make
   ./data_preprocessing  
   

Python Environment Setup

1. Install Python Dependencies:

   cd training
   pip install -r requirements.txt

2. Required Libraries:

   • torch
   • torch-geometric
   • pandas
   • matplotlib
   • scipy
   • networkx
   • plotly

---

Running the Project

1. Data Preprocessing (C++):

Navigate to the src directory and run the data preprocessing script.

cd src/build
./data_preprocessing

This will generate filtered datasets in the data/filtered/ directory using the data/unfiltered/ directory. You might need to manually paste the data from Kaggle to data/unfiltered/, due to size limitations on Github.

2. Training the GNN Model (Python):

Navigate to the training directory and run the training.ipynb notebook.

cd training
jupyter notebook training.ipynb

This will train the GNN model and save the model weights to gcn_model_weights.pth.

3. Visualizing the Results (Python):

Navigate to the visualization directory and run the data_visualization.ipynb notebook.

cd visualization
jupyter notebook data_visualization.ipynb

Visualizations

Data Plot

Data Predictions Plot

Usage

• Training: The GNN model can be trained using the training.ipynb notebook. Adjust hyperparameters as needed within the notebook.
• Visualization: Use the data_visualization.ipynb notebook to generate visualizations of the transaction graph and model performance metrics.

Conclusion

In this model, each node aggregates information from its first-order neighbors in both GCN layers. Although the second GCN layer also considers first-order neighbors, these neighbors' features have already been influenced by their own neighbors in the previous layer. This way, each node indirectly incorporates second-order neighbor information as well. However, the direct aggregation occurs only from first-order neighbors in each GCN layer.

License

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

Contributing

Contributions are welcome! Please fork the repository and submit a pull request for any enhancements or bug fixes.

Acknowledgments

• The Elliptic dataset: Kaggle
• PyTorch Geometric: PyTorch Geometric

Contact

For any questions or suggestions, please open an issue or contact the project maintainers.