lts4 / gnn_statistics

Code for the paper "On the choice of graph neural network architectures"

On the choice of graph neural network architectures

Clément Vignac, Guillermo Ortiz-Jiménez, Pascal Frossard, On the choice of graph neural network architectures, ICASSP 2020

TensorFlow based benchmarking of graph neural networks on citation data.

This code is based on the GNN-benchmark library.

Installation

Follow the instructions of GNN-benchmark for installing the library.

Running experiments

Don't forget to turn MongoDB on before running the code.

Experiment 1a.

In this experiment we compare the performance of Graph Convolutional Networks and Simple Graph Convolutions on Cora, Pubmed and Citeseer when 50% of the nodes are observed. Different numbers of random features are used.

#!/usr/bin/env bash
python3 scripts/create_jobs.py -c config/experiment1a.conf.yaml --op search
python3 scripts/spawn_worker.py -c config/experiment1a.conf.yaml --gpu 0
python3 scripts/aggregate_results.py -c config/experiment1a.conf.yaml -o results/experiment1a/

To check that the code runs, you can reduce the number of splits in config/experiment1a.conf.yaml.

Experiment 1b.

In this experiment the number of features is fixed (at 300), but the proportion of observed nodes varies.

#!/usr/bin/env bash
python3 scripts/create_jobs.py -c config/experiment1b.conf.yaml --op search
python3 scripts/spawn_worker.py -c config/experiment1b.conf.yaml --gpu 0

Experiment 2

With 50% of observed nodes and 300 features, compare the performance of GCN, SGC, APPNP as well as non-linear version of them (APPNP and SGC).

#!/usr/bin/env bash
python3 scripts/create_jobs.py -c config/experiment2.conf.yaml --op fixed
python3 scripts/spawn_worker.py -c config/experiment2.conf.yaml --gpu 0
mkdir results/experiment2
python3 scripts/aggregate_results.py -c config/experiment2.conf.yaml -o results/experiment2/

Cite

Please cite our paper as well as GNN-benchmark if you use this code:

@article{vignac2019choice,
  title={On the choice of graph neural network architectures},
  author={Vignac, Cl{\'e}ment and Ortiz-Jim{\'e}nez, Guillermo and Frossard, Pascal},
  journal={arXiv preprint arXiv:1911.05384},
  year={2019}
}

@article{shchur2018pitfalls,
  title={Pitfalls of Graph Neural Network Evaluation},
  author={Shchur, Oleksandr and Mumme, Maximilian and Bojchevski, Aleksandar and G{\"u}nnemann, Stephan},
  journal={Relational Representation Learning Workshop, NeurIPS 2018},
  year={2018}
}