switiz/gnn-gcn-gat

example of gnns

GNN(Graph Neural Network)

Provide Example of Graph Neural Networks

---

index

• Benchmark Dataset
• GCN(Graph Convolutional Network)
• GAT(Graph Attention Network)

---

Benchmark Dataset

Cora

sota : https://paperswithcode.com/sota/node-classification-on-cora

Citeseer

sota : https://paperswithcode.com/sota/node-classification-on-citeseer

Pubmed

sota : https://paperswithcode.com/sota/node-classification-on-pubmed

---

GCN (Graph Convolutional Networks)

https://arxiv.org/abs/1609.02907

Overview

---

GAT (Graph Attention Network)

https://openreview.net/forum?id=rJXMpikCZ

Overview

---

Train

usage : 
$ python train.py --epochs 1000

mode:
- gat
- spgat
- gcn

dataset :
- cora
- citeseer
- pubmed (not tested)

parameters:

--model default='gat', help='select model'
--data default='cora', help='select dataset'
--epochs default=100000, help='Number of epochs to train.'
--niter default=10, help='iter value for avg.'
--early_stopping default=False, help='set early_stopping'
--lr default=0.005, help='Initial learning rate.'
--seed default=72, help='Random seed.'
--weight_decay default=5e-4, help='Weight decay (L2 loss on parameters).'
--hidden default=8, help='Number of hidden units.'
--nb_heads default=8, help='Number of head attentions.'
--dropout default=0.6, help='Dropout rate (1 - keep probability).'
--alpha default=0.2, help='Alpha for the leaky_relu.'
--patience default=100, help='Patience'
--verbose default=False, help='set early_stopping'

parameters usage:
$ python train.py --epochs 1000 --data citeseer --verbose --seed 80

Result

	GCN	GAT	SPGAT
Cora	0.816	0.814	0.812
Citeseer	0.702	0.7039	0.7083

Speed: GCN > SPGAT > GAT

Reference

@article{
  velickovic2018graph,
  title="{Graph Attention Networks}",
  author={Veli{\v{c}}kovi{\'{c}}, Petar and Cucurull, Guillem and Casanova, Arantxa and Romero, Adriana and Li{\`{o}}, Pietro and Bengio, Yoshua},
  journal={International Conference on Learning Representations},
  year={2018},
  url={https://openreview.net/forum?id=rJXMpikCZ},
  note={accepted as poster},
}
@article{kipf2016semi,
  title={Semi-Supervised Classification with Graph Convolutional Networks},
  author={Kipf, Thomas N and Welling, Max},
  journal={arXiv preprint arXiv:1609.02907},
  year={2016}
}
@inproceedings{nr,
     title={The Network Data Repository with Interactive Graph Analytics and Visualization},
     author={Ryan A. Rossi and Nesreen K. Ahmed},
     booktitle={AAAI},
     url={http://networkrepository.com},
     year={2015}
}

GCN

[Tensorflow] https://github.com/tkipf/gcn
[Pytorch] https://github.com/tkipf/pygcn
[Pytorch] https://github.com/marblet/gcnns

GAT

[Tensorflow] https://github.com/PetarV-/GAT
[Keras] https://github.com/danielegrattarola/keras-gat
[Pytorch] https://github.com/marblet/gat-pytorch
[Pytorch] https://github.com/Diego999/pyGAT