Origin

This repository was created as a result of a class project in CSE 6521 instructed by Dr. Yu Su at Ohio State University. Ronald Davies and myself (David Carlyn) are the authors of this code and report given. See the section below for the original work and authors that this code is based on.

Attempted Reimplementation of the iVGAE model:

Based on the interventional variational graph autoencoder from the paper Relating Graph Neural Networks to Structural Causal Models

Authors: Matej Zecevic, Devendra Singh Dhami, Petar Velickovic, Kristian Dersting

Packages

Pytorch 1.12.0
conda install pytorch==1.12.1 torchvision==0.13.1 torchaudio==0.12.1 cudatoolkit=11.6 -c pytorch -c conda-forge

Pytorch Geometric
conda install pyg -c pyg

Matplotlib
conda install matplotlib

Base start from PytorchGeometric tutorial:
https://colab.research.google.com/github/AntonioLonga/PytorchGeometricTutorial/blob/main/Tutorial6/Tutorial6.ipynb

iVGAE Test Run

The base model can be tested on the Citeseer dataset by running:
python main.py

Training without interventions

The base model can be trained on either the ASIA (asia) or the Earthquake (earthquake) datasets:
python train.py --dset asia --epochs 200

Training with interventions

To train with an intervention, provide a list of variables to intervene on and their probabilities:
python train.py --dset asia --epochs 200 --i_vars T --i_probs 0.50

Testing without interventions

python test.py --model [PATH_TO_MODEL] --dset asia --num_of_runs 50

Testing with interventions

python test.py --model [PATH_TO_MODEL] --dset asia --num_of_runs 50 --i_vars T --i_probs 0.50