Skip to content

Latest commit

 

History

History
49 lines (40 loc) · 1.68 KB

README.md

File metadata and controls

49 lines (40 loc) · 1.68 KB
Raw

Fraud Detection with Graph Neural Networks

Detect fraud in transactions graphs using various ready-to-use models and datasets.

Models:

  • GAT
  • GCN
  • GIN
  • GraphSAGE
  • MPNN
  • GTN

Datasets:

Installation

First install the requirements.

pip install -r requirements.txt

Download Dataset

Download and extract the required datasets. This project currently only supports Elliptic. A dataset must have 3 files:

  • Edge list
  • Classes
  • Features

Then provide the required torch.utils.data.Dataset class for your own dataset like here

Training

To train a model you need a config file. By default, these files are located at configs/. For example to train a GAT model on Elliptic run:

python train --config configs/elliptic_gat.yaml

The models are trained for 100 epochs and results are verbosed every 10 epochs and logged to Tensorboard.

Visualization

To visualize model predictions on the graph on each step, run:

python visualize.py --config configs/elliptic_gat.yaml --step 30 --weights_file weigths/elliptic_gat.pt

A sample predictions visualization

About the project

This is my final project for the course Complex Networks during my Masters' in Shahid Beheshti University instructed by Dr. Sadegh Ali Akbari (Spring 2023). I've also written a Persian blog post on this project here