Bag of Tricks for Node Classification with Graph Neural Networks

The official implementation for Bag of Tricks for Node Classification with Graph Neural Networks (Best Paper Award at DLG-KDD'21 workshop) based on Deep Graph Library.

Dependencies

• dgl 0.5.*
• torch 1.6.0
• ogb 1.3.1

How to run

Cora, Citeseer, Pubmed, Reddit, ogbn-arxiv

Run

cd src/no-sampling/
python3 run.py [args]

For example,

python3 run.py --optimizer=rmsprop --lr=0.002 --loss=loge --labels --mask-rate=0.5 --model=gat --linear --n-heads=3 --n-hidden=250 --dropout=0.75 --input-drop=0.25 --attn-drop=0.1 --norm-adj=symm

More details of the hyperparameters and experimental results can be found at the end of run.py.

ogbn-proteins

Run

cd src/ogbn-proteins/
python3 gat.py [args]

For the results in the paper, run

python3 gat.py

or

python3 gat.py --use-labels

ogbn-products

First change the directory

cd src/ogbn-products/

For GAT, run

python3 gat.py [args]

For MLP, run

python3 mlp.py [args]

Citing our work

If you find this work helpful in your research, please consider citing our work.

@article{wang2021bag,
  title={Bag of Tricks for Node Classification with Graph Neural Networks},
  author={Wang, Yangkun and Jin, Jiarui and Zhang, Weinan and Yu, Yong and Zhang, Zheng and Wipf, David},
  journal={arXiv preprint arXiv:2103.13355},
  year={2021}
}