The goal of this work is to tackle anti-money laundering problem, trying to classify efficiently illicit transactions and create a comparison between different architectures, in particular comparing different types of Graph Neural Networks (GNNs), to identify what are the key features and approaches that enable good performances in this given context.
First of all, you have to clone the repository with the standard command:
git clone https://github.com/simonemarasi/aml-elliptic-gnn
If you want only run the code note that it is also available a ready-to-run Google Colab version of the project at the following link. The full code is inspectable, cloning this repository.
You can download the dataset zipped from the following link. After done that make sure to extract the zip file into the data folder located at the root of the repository. However, you can put the data also in other places, making sure to change the folder in the configuration file (config.yaml) accordingly.
In the configuration file, it is possible also to modify some hyperparameters such as the number of epochs, the number of hidden units to use, the learning rate, etc.
It is possible to install all the packages required for the execution launching the command
pip install -r requirements.txt
After that, to run the script execute the command
python main.py
| Model | Precision | Recall | F1 | F1 Micro AVG |
|---|---|---|---|---|
| GCN | 0.832 | 0.457 | 0.59 | 0.94 |
| GAT | 0.787 | 0.683 | 0.731 | 0.952 |
| SAGE | 0.931 | 0.788 | 0.853 | 0.974 |
| Cheb | 0.942 | 0.795 | 0.862 | 0.976 |
| GATv2 | 0.891 | 0.804 | 0.845 | 0.972 |
| Custom GAT | 0.861 | 0.762 | 0.808 | 0.966 |
You can find also the complete report of the project in this repository.
The paper using this work was presented at 2024 IEEE 21st Consumer Communications & Networking Conference (CCNC) held in Las Vagas on January 2024.