Graphtester

Graphtester is a Python package for comprehensive analysis of the theoretical capabilities of GNNs for various datasets, tasks, and scores, using an extension of the Weisfeiler-Lehman framework.

You can read more about the theoretical background in our paper.

If you use Graphtester in your research, please cite:

@inproceedings{akbiyik2023graphtester,
    title        = {{Graphtester: Exploring Theoretical Boundaries of GNNs on Graph Datasets}},
    author       = {Eren Akbiyik and Florian Grötschla and Béni Egressy and Roger Wattenhofer},
    year         = 2023,
    month        = {July},
    booktitle    = {{Data-centric Machine Learning Research (DMLR) Workshop at ICML 2023, Honolulu, Hawaii}}
}

See documentation at https://graphtester.readthedocs.io.

Reproducing the results in the paper

1. Install the package
2. Run sbatch scripts/paper/test_all_datasets.sh to run all experiments on the datasets. Note that this scripts requires 10 cores and substantial amount of memory. You can change the number of cores by changing the --cpus-per-task argument in the script.
3. Recommendation results can be gathered by adjusting the parameters in scripts/paper/test_all_datasets.py script for the desired task and running the bash script again.

Development

1. Install poetry
2. Run poetry install to install dependencies
3. Run poetry run pre-commit install to install pre-commit hooks

Environment

You might want to set the following environment variables:

export PYTHON_KEYRING_BACKEND=keyring.backends.null.Keyring # disable keyring, might be needed for poetry
export SCRATCH=/itet-stor/${USER}/net_scratch # set cache locations
export XDG_CACHE_HOME=$SCRATCH/.cache # generic Linux cache
export POETRY_VIRTUALENVS_PATH=$SCRATCH/envs # poetry environment
export GRAPHTESTER_CACHE_DIR=$SCRATCH/.graphtester # graphtester cache