A Fair Comparison of Graph Neural Networks for Graph Classification (ICLR 2020)

Summary

The library includes data and scripts to reproduce the experiments reported in the paper.

If you happen to use or modify this code, please remember to cite our paper:

Federico Errica, Marco Podda, Davide Bacciu, Alessio Micheli: A Fair Comparison of Graph Neural Networks for Graph Classification. Proceedings of the 8th International Conference on Learning Representations (ICLR 2020).

@inproceedings{errica_fair_2020,
    title = {A fair comparison of graph neural networks for graph classification},
    booktitle = {Proceedings of the 8th {International} {Conference} on {Learning} {Representations} ({ICLR})},
    author = {Errica, Federico and Podda, Marco and Bacciu, Davide and Micheli, Alessio},
    year = {2020}
}

--

Updated Table with Results (CHEMICAL)

	D&D	NCI1	PROTEINS
Baseline	$\mathbf{78.4}\pm 4.5 $	$69.8 \pm 2.2 $	$\mathbf{75.8} \pm 3.7 $
DGCNN	$76.6 \pm 4.3 $	$76.4 \pm 1.7 $	$72.9 \pm 3.5 $
DiffPool	$75.0 \pm 3.5 $	$76.9 \pm 1.9 $	$73.7 \pm 3.5 $
ECC	$72.6 \pm 4.1 $	$76.2 \pm 1.4 $	$72.3 \pm 3.4 $
GIN	$75.3 \pm 2.9 $	$\mathbf{80.0} \pm 1.4 $	$73.3 \pm 4.0 $
GraphSAGE	$72.9 \pm 2.0 $	$76.0 \pm 1.8 $	$73.0 \pm 4.5 $
CGMM	$74.9 \pm 3.4 $	$76.2 \pm 2.0$	$74.0 \pm 3.9$
ECGMM	$73.9 \pm4.1$	$78.5 \pm 1.7$	$73.3 \pm 4.1$
iCGMMf	$75.1 \pm 3.8$	$76.4 \pm1.4$	$73.2 \pm 3.9$
GSPN	-	$76.6 \pm 1.9$	-

Updated Table with Results (SOCIAL + degree)

	IMDB-B	IMDB-M	REDDIT-B	REDDIT-5K	COLLAB
Baseline	$70.8 \pm 5.0 $	$\mathbf{49.1} \pm 3.5 $	$82.2 \pm 3.0 $	$52.2 \pm 1.5 $	$70.2 \pm 1.5 $
DGCNN	$69.2 \pm 3.0 $	$45.6 \pm 3.4 $	$87.8 \pm 2.5 $	$49.2 \pm 1.2 $	$71.2 \pm 1.9 $
DiffPool	$68.4 \pm 3.3 $	$45.6 \pm 3.4 $	$89.1 \pm 1.6 $	$53.8 \pm 1.4 $	$68.9 \pm 2.0 $
ECC	$67.7 \pm 2.8 $	$43.5 \pm 3.1 $	-	-	-
GIN	$71.2 \pm 3.9 $	$48.5 \pm 3.3 $	$89.9 \pm 1.9 $	$\mathbf{56.1} \pm 1.7 $	$75.6 \pm 2.3 $
GraphSAGE	$68.8 \pm 4.5 $	$47.6 \pm 3.5 $	$84.3 \pm 1.9 $	$50.0 \pm 1.3 $	$73.9 \pm 1.7 $
CGMM	$\mathbf{72.7} \pm 3.6$	$47.5 \pm 3.9$	$88.1 \pm 1.9$	$52.4 \pm 2.2$	$77.32 \pm 2.2$
ECGMM	$70.7 \pm 3.8$	$48.3 \pm 4.1 $	$89.5 \pm 1.3$	$53.7 \pm 1.0 $	$77.45 \pm 2.3$
iCGMMf	$71.8 \pm 4.4$	$49.0 \pm 3.8 $	$\mathbf{91.6} \pm 2.1$	$55.6 \pm 1.7$	$\mathbf{78.9} \pm 1.7$
GSPN	-	-	$90.5 \pm 1.1$	$55.3 \pm 2.0$	$78.1 \pm 2.5$

If you are interested in an introduction to Deep Graph Networks (and a new library!), check this out:

Bacciu Davide, Errica Federico, Micheli Alessio, Podda Marco: A Gentle Introduction to Deep Learning for Graphs, Neural Networks, 2020. DOI: 10.1016/j.neunet.2020.06.006.

Installation

We provide a script to install a virtual environment called gnn-comparison. You will a Python version installed on your system.

To install the required packages, follow there instructions (tested on a linux terminal):

1. clone the repository

   git clone https://github.com/diningphil/gnn-comparison

2. cd into the cloned directory

   cd gnn-comparison

3. change the PYTHON_VERSION variable in install.sh to your system's Python version.

If you want to recreate the original environment used for the paper:

4. run the installation script

   source install_original.sh [<your_cuda_version>]
   

Where <your_cuda_version> is an optional argument that can be either cpu, cu92, cu100 or cu101.

Pytorch Geometric 1.4.0 will also be installed.

Otherwise, for newer versions

4. run the installation script

   source install.sh [<your_cuda_version>] [<your_pytorch_version>]
   

Where <your_pytorch_version> should be >= 2.0.1, and <your_cuda_version> is an optional argument that can be either cpu, cu116, cu117 or cu118. If you do not provide any of them the script will default to Pytorch 2.0.1 and cpu.

Pytorch Geometric 2.3.1 will also be installed.

Notes:

• It is up to you to ensure the Python version is consistent with the Pytorch, Torch Geometric, and CUDA versions you are going to install
• Please open an issue if something is not working as expected.

Running the experiments

To reproduce the experiments, first preprocess datasets as follows:

python PrepareDatasets.py DATA/CHEMICAL --dataset-name <name> --outer-k 10

python PrepareDatasets.py DATA/SOCIAL_1 --dataset-name <name> --use-one --outer-k 10

python PrepareDatasets.py DATA/SOCIAL_DEGREE --dataset-name <name> --use-degree --outer-k 10

Where <name> is the name of the dataset. Then, substitute the split (json) files with the ones provided in the data_splits folder.

Please note that dataset folders should be organized as follows:

CHEMICAL:
    NCI1
    DD
    ENZYMES
    PROTEINS
SOCIAL[_1 | _DEGREE]:
    IMDB-BINARY
    IMDB-MULTI
    REDDIT-BINARY
    REDDIT-MULTI-5K
    COLLAB

Then, you can launch experiments by typing:

cp -r DATA/[CHEMICAL|SOCIAL_1|SOCIAL_DEGREE]/<name> DATA
python Launch_Experiments.py --config-file <config> --dataset-name <name> --result-folder <your-result-folder> --debug

Where <config> is your config file (e.g. config_BaselineChemical.yml), and <name> is the dataset name chosen as before.

Additional Notes

You can only use CUDA with the --debug option, parallel GPUs support is not provided.

Troubleshooting

If you would like PyTorch not to spawn multiple threads for each process (highly recommended), append export OMP_NUM_THREADS=1 to your .bashrc file.