Paper links: Short paper | Full paper
Implementation for small graph datasets: this repo
Implementation for LINKX large graph datasets : repo
Experiments were conducted with following setup:
Pytorch: 1.6.0
Python: 3.8.5
Cuda: 10.2.89
Trained on NVIDIA V100 GPU.
Main Experiment Settings:
Hidden dimension: 64
No. of hops: 3 ( Total 7 feature matrices )
p : 4 ( up to 4 feature matrices can be selected )
Summary of results
| Dataset | Accuracy (%) | Avg. train time, 3-runs (sec) |
|---|---|---|
| Cora | 87.77 | 36.53 |
| Citeseer | 77.15 | 46.64 |
| Pubmed | 89.64 | 71.58 |
| Chameleon | 78.46 | 59.95 |
| Wisconsin | 89.41 | 19.45 |
| Texas | 87.57 | 46.83 |
| Cornell | 88.11 | 44.84 |
| Squirrel | 73.97 | 88.55 |
| Actor | 37.29 | 52.13 |
Run node classification:
./run_classification.sh
(Results may vary slightly with a different platform, e.g. use of different GPU. In such case, for best performance, some hyperparameter search may be required. Please refer to PyTorch documentation for more details.)
Datasets and parts of preprocessing code were taken from Geom-GCN and GCNII repositories. We thank the authors of these papers for sharing their code.