SigMaNet: One Laplacian to Rule Them All

This repository is the offical PyTorch implementation of SigMaNet: both its implementation and code for running other convolutional graph networks is present.

Enviroment Setup

The experiments were conducted under this specific environment:

1. Ubuntu 20.04.3 LTS
2. Python 3.8.10
3. CUDA 10.2
4. Torch 1.11.0 (with CUDA 10.2)

In addition, torch-scatter, torch-sparse and torch-geometric are needed to handle scattered graphs and train Graph Neural Networks (GNNs) for a wide range of applications related to structured data. For these three packages, follow the official instructions for torch-scatter, torch-sparse, and torch-geometric.

Finally, Pytorch Geometric Signed Directed GitHub Pages must be installed. Other library are listed below

Repository structure

The repository contains two folders:

• data contains the syntactic graphs in synthetic and the WikiRfa dataset in wikirfa.
• src contains all the model implementations used for running the experiments. Futhermore, it stores two other foldes utils and layer.

Run code

cd src
python3 node_SigMaNet.py --dataset dataset_nodes500_alpha0.05_beta0.2
python3 Edge_SigMaNet.py --dataset dataset_nodes500_alpha0.05_beta0.2 --task direction --noisy -N

License

SigMaNet is released under the MIT License

Acknowledgements

The template is borrowed from MagNet and Pytorch-Geometric Signed Directed. We thank the authors for the excellent repositories.