dnx's Introduction

DnX and FastDnX

This is repo for Distill n' Explain: explaining graph neural networks using simple surrogates. This folder contains the source code of using DnX and FastDnX to explain node predictions.

Authors:

Requirements

We have built this implementation using:

PyTorch
PyTorch Geometric

Folder description:

trained_gcn: contains the models that will be explained
trained_distiller: contains the distilled models
explanation: contains the explanations generated by our models

file description

model.py: the model used to distill.
utils.py: all functions used to load, save and evaluate the dataset
main_distiller.py: code with the functions to distill the models to be explained
main_syn.ipynb: notebook to perform the task of generating explanations for the synthetic datasets

Datasets:

The nomenclature (syn1, syn2 ...., syn6) is used to facilitate experiments

BA-HouseShapes (syn1),
BA-Community (syn2),
BA-Grids (syn3),
Tree-Cycles (syn4),
TreeGrids (syn5),
BA-Bottle-Shaped (syn6)

To run the DnX and Fast DnX explainers for the synthetic datasets just run the notebook main_syn.ipynb:

Cite

@inproceedings{distillnexplain2023,
  title={Distill n' Explain: explaining graph neural networks using simple surrogates},
  author={T. A. Pereira and E. J. Nascimento and L. E. Resck and D. Mesquita and A. H. Souza},
  booktitle={Artificial Intelligence and Statistics (AISTATS)},
  year={2023}
}

Recommend Projects