Awesome

Official Pytorch Implementation of "AdaGCN: Adaboosting Graph Convolutional Networks into Deep Models" (ICLR 2021)

Please refer to openreview (ICLR 2021) to look into the details of our paper.

Enviromment

python3.6  
cuda11.0  
torch1.7.1

Run the code （Datasets: citeseer, cora_ml, pubmed and ms_academic）

Baseline: GCN

python main.py --trainsize 20 --dataset citeseer --niter 5 --nseed 20 --model GCN --dropout 0.5 --reg 5e-4

Baseline: APPNP or PPNP

python main.py --trainsize 20 --dataset citeseer --niter 5 --nseed 20 --model APPNP --dropout 0.5 --early 1 --patience 300 --max 500 --reg 5e-3

AdaGCN on Four datasets:

python main.py --trainsize 20 --dataset citeseer --niter 5 --nseed 20 --model AdaGCN --layers 15 --hid_AdaGCN 5000 --dropout 0.0 --weight_decay 1e-3 --early 1 --patience 300 --max 500 --reg 5e-3   
python main.py --trainsize 20 --dataset cora_ml --niter 5 --nseed 20 --model AdaGCN --layers 12 --hid_AdaGCN 5000 --dropout 0.0 --weight_decay 1e-4 --early 1 --patience 300 --max 500 --reg 5e-3   
python main.py --trainsize 20 --dataset pubmed --niter 5 --nseed 20 --model AdaGCN --layers 20 --hid_AdaGCN 5000 --dropout 0.2 --weight_decay 1e-4 --early 1 --patience 300 --max 500 --reg 5e-3   
python main.py --trainsize 20 --dataset ms_academic --niter 5 --nseed 20 --model AdaGCN --layers 5 --hid_AdaGCN 3000 --dropout 0.2 --weight_decay 1e-4 --early 1 --patience 300 --max 500 --reg 5e-3

Results:

Acknowledgement

Our code is directly adapted from PPNP paper Predict then Propagate: Graph Neural Networks meet Personalized PageRank (ICLR 2019) github: https://github.com/klicperajo/ppnp.

Contact

Please refer to ajksunke@pku.edu.cn in case you have any questions.

Cite

Please cite our paper if you use the model or this code in your own work:

@inproceedings{sun2020adagcn,
  title={AdaGCN: Adaboosting Graph Convolutional Networks into Deep Models},
  author={Sun, Ke and Zhu, Zhanxing and Lin, Zhouchen},
  booktitle={International Conference on Learning Representations},
  year={2020}
}