An Experimental Study of the Transferability of Spectral Graph Networks

This repository holds the corresponding code to the paper ´An Experimental Study of the Transferability of Spectral Graph Networks´ by Axel Nilsson and Xavier Bresson.

The focus of the work is to study the performance of the ChebNet, a spectral graph neural network, with regards to other spacial methods on datasets made out of sets of graphs. The datasets are all from the open benchmarks OGB and benchmarking-gnns. The figure below shows the summary of the chosen tasks.

Benchmarking gnns

Most of the structure of the code are simplifications made on the benchmarking-gnns code base

Open Graph Benchmark (OGB)

The code used for OGB is straightforward with small changes to the original example notebooks. We use the DGL framework to make sure to use the same model as for the benchmarking gnns experiments.

To reproduce the results:

Setup your environment:

conda env create -f environment_gpu.yml

OGB:

In order to run a model on a given dataset $DATASET with an output file $FILENAME run the following command:
```
  python main_dgl.py --dataset $DATASET --gnn Cheb_net --filename $FILENAME
```
Otherwise make sure to make a script that matches your config by tweaking on the script script_ogb.sh.

Benchmarking_gnns:

In order to make run the benchmarking gnn datasets you should first download them by using the following commands:
```
  cd Benchmark-gnn/data/
  bash script_download_all_datasets.sh
```
Then you can run the experiments by using the command:
```
  bash benchmark_gnn_script.sh
```

Repository structure

.
├── Benchmark-gnn/
    ├── configs/             # Files containing the configuration for the models for each task
    ├── data/ 
        ├──  Molecules
        └──  SBMS
    ├── layers/              # Definition of the ChebNet layer
    ├── nets/                # Definition of the structure of the NNs for each model
    ├── train/               # Trainig script for each task
    └── scripts              # Sets of scripts to run individually each task
├── OGB/
    ├── Dataset/
    ├── gnn_dgl.py           # Model definiton
    └── main_dgl.py          # Model training anf testing script
├── environment files        # Yml files used to set up the environment
├── LICENSE
└── README.md

Results

The results as can be found in the paper:

Reference

arXiv's paper

@misc{nilsson2020experimental,
      title={An Experimental Study of the Transferability of Spectral Graph Networks}, 
      author={Axel Nilsson and Xavier Bresson},
      year={2020},
      eprint={2012.10258},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Axeln78/Transferability-of-spectral-gnns