cenyk1230/crgnn

This repository is an implementation of paper : Improving the Training of Graph Neural Networks with Consistency Regularization

CRGNN

Paper ： Improving the Training of Graph Neural Networks with Consistency Regularization

Environments

Implementing environment: GeForce RTX™ 3090 24GB (GPU)

Requirements

pytorch>=1.8.1

ogb=1.3.2

numpy=1.21.2

cogdl (latest version)

Training

GAMLP+RLU+SCR

For ogbn-products:

Params: 3335831

python pre_processing.py --num_hops 5 --dataset ogbn-products

python main.py --use-rlu --method R_GAMLP_RLU --stages 400 300 300 300 300 300 --train-num-epochs 0 0 0 0 0 0 --threshold 0.85 --input-drop 0.2 --att-drop 0.5 --label-drop 0 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --act leaky_relu --batch_size 50000 --patience 300 --n-layers-1 4 --n-layers-2 4 --bns --gama 0.1 --consis --tem 0.5 --lam 0.1 --hidden 512 --ema

GAMLP+MCR

For ogbn-products:

Params: 3335831

python pre_processing.py --num_hops 5 --dataset ogbn-products

python main.py --use-rlu --method R_GAMLP_RLU --stages 800 --train-num-epochs 0 --input-drop 0.2 --att-drop 0.5 --label-drop 0 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --act leaky_relu --batch_size 100000 --patience 300 --n-layers-1 4 --n-layers-2 4 --bns --gama 0.1 --tem 0.5 --lam 0.5 --ema --mean_teacher --ema_decay 0.999 --lr 0.001 --adap --gap 10 --warm_up 150 --top 0.9 --down 0.8 --kl --kl_lam 0.2 --hidden 512

GIANT-XRT+GAMLP+MCR

Please follow the instruction in GIANT to get the GIANT-XRT node features.

For ogbn-products:

Params: 2144151

python pre_processing.py --num_hops 5 --dataset ogbn-products --giant_path " "

python main.py --use-rlu --method R_GAMLP_RLU --stages 800 --train-num-epochs 0 --input-drop 0.2 --att-drop 0.5 --label-drop 0 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --act leaky_relu --batch_size 100000 --patience 300 --n-layers-1 4 --n-layers-2 4 --bns --gama 0.1 --tem 0.5 --lam 0.5 --ema --mean_teacher --ema_decay 0.99 --lr 0.001 --adap --gap 10 --warm_up 150 --kl --kl_lam 0.2 --hidden 256 --down 0.7 --top 0.9 --giant

SAGN+MCR

For ogbn-products:

Params: 2179678

python pre_processing.py --num_hops 3 --dataset ogbn-products

python main.py --method SAGN --stages 1000 --train-num-epochs 0 --input-drop 0.2 --att-drop 0.4 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --batch_size 100000 --patience 300 --tem 0.5 --lam 0.5 --ema --mean_teacher --ema_decay 0.99 --lr 0.001 --adap --gap 20 --warm_up 150 --top 0.85 --down 0.75 --kl --kl_lam 0.01 --hidden 512 --zero-inits --dropout 0.5 --num-heads 1  --label-drop 0.5  --mlp-layer 2 --num_hops 3 --label_num_hops 14

GIANT-XRT+SAGN+MCR

Please follow the instruction in GIANT to get the GIANT-XRT node features.

For ogbn-products:

Params: 1154654

python pre_processing.py --num_hops 3 --dataset ogbn-products --giant_path " "

python main.py --method SAGN --stages 1000 --train-num-epochs 0 --input-drop 0.2 --att-drop 0.4 --pre-process --residual --dataset ogbn-products --num-runs 10 --eval 10 --batch_size 50000 --patience 300 --tem 0.5 --lam 0.5 --ema --mean_teacher --ema_decay 0.99 --lr 0.001 --adap --gap 20 --warm_up 100 --top 0.85 --down 0.75 --kl --kl_lam 0.02 --hidden 256 --zero-inits --dropout 0.5 --num-heads 1  --label-drop 0.5  --mlp-layer 1 --num_hops 3 --label_num_hops 9 --giant

Use Optuna to search for C&S hyperparameters

We searched hyperparameters using Optuna on validation set.

python post_processing.py --file_name --search

GAMLP+RLU+SCR+C&S

python post_processing.py --file_name --correction_alpha 0.4780826957236622 --smoothing_alpha 0.40049734940262954

GIANT-XRT+SAGN+MCR+C&S

python post_processing.py --file_name --correction_alpha 0.42299283241438157 --smoothing_alpha 0.4294212449832242

Node Classification Results:

Performance on ogbn-products(10 runs):

Methods	Validation accuracy	Test accuracy
SAGN+MCR	0.9325±0.0004	0.8441±0.0005
GAMLP+MCR	0.9319±0.0003	0.8462±0.0003
GAMLP+RLU+SCR	0.9292±0.0005	0.8505±0.0009
GAMLP+RLU+SCR+C&S	0.9304±0.0005	0.8520±0.0008
GIANT-XRT+GAMLP+MCR	0.9402±0.0004	0.8591±0.0008
GIANT-XRT+SAGN+MCR	0.9389±0.0002	0.8651±0.0009
GIANT-XRT+SAGN+MCR+C&S	0.9387±0.0002	0.8673±0.0008

Citation

Our paper:

@misc{zhang2021improving,
      title={Improving the Training of Graph Neural Networks with Consistency Regularization}, 
      author={Chenhui Zhang and Yufei He and Yukuo Cen and Zhenyu Hou and Jie Tang},
      year={2021},
      eprint={2112.04319},
      archivePrefix={arXiv},
      primaryClass={cs.SI}
}

GIANT paper:

@article{chien2021node,
  title={Node Feature Extraction by Self-Supervised Multi-scale Neighborhood Prediction},
  author={Eli Chien and Wei-Cheng Chang and Cho-Jui Hsieh and Hsiang-Fu Yu and Jiong Zhang and Olgica Milenkovic and Inderjit S Dhillon},
  journal={arXiv preprint arXiv:2111.00064},
  year={2021}
}

GAMLP paper:

@article{zhang2021graph,
  title={Graph attention multi-layer perceptron},
  author={Zhang, Wentao and Yin, Ziqi and Sheng, Zeang and Ouyang, Wen and Li, Xiaosen and Tao, Yangyu and Yang, Zhi and Cui, Bin},
  journal={arXiv preprint arXiv:2108.10097},
  year={2021}
}

SAGN paper:

@article{sun2021scalable,
  title={Scalable and Adaptive Graph Neural Networks with Self-Label-Enhanced training},
  author={Sun, Chuxiong and Wu, Guoshi},
  journal={arXiv preprint arXiv:2104.09376},
  year={2021}
}

C&S paper:

@inproceedings{
huang2021combining,
title={Combining Label Propagation and Simple Models out-performs Graph Neural Networks},
author={Qian Huang and Horace He and Abhay Singh and Ser-Nam Lim and Austin Benson},
booktitle={International Conference on Learning Representations},
year={2021},
url={https://openreview.net/forum?id=8E1-f3VhX1o}
}