Systematic relational reasoning using Epistemic Graph Neural Networks

Getting started

We offer this project as dev package that can be built in editable mode:

Create a virtual environment and install the necessary packages below

python -m venv venv
source venv/bin/activate # for linux
.\venv\Scripts\activate # for windows
pip install -e .

Also install the following packages using the link relevant for your hardware:

pip install torch-scatter torch-sparse -f https://data.pyg.org/whl/torch-2.1.0+${CUDA}.html

e.g. CUDA=cpu for a machine without a gpu. See https://pypi.org/project/torch-sparse/

To reproduce results in the paper just build and run the following command in src

python train.py experiments=`pick a dataset model config file from configs/experiments`

The config can be easily adjusted from the command line using a dot file path notation with the experiments. prefix. For example, running for 10 epochs on the rcc8 dataset amounts to:

python train.py experiments=fb_model_rcc8 experiments.epochs=10

Find STaR on Huggingface datasets

You can also find the disjunctive systematic relational generalization STaR dataset that was developed along with this work on huggingface

from datasets import load_dataset

star_dataset = load_dataset("erg0dic/STaR")

Cite

If you find this code useful, please consider citing us:

@inproceedings{
khalid2025systematic,
title={Systematic Relational Reasoning With Epistemic Graph Neural Networks},
author={Irtaza Khalid and Steven Schockaert},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025},
url={https://openreview.net/forum?id=qNp86ByQlN}
}