A Biased Graph Neural Network Sampler with Near-Optimal Regret

The implementation and expremental code of Thanos graph neural network sampler See paper.

Datasets

• Chameleon
• Squirrel
• CoraFull
• ogbn-arxiv
• ogbn-product

Model

• GCN
• GAT

Dependencies

• tensorflow==1.13.1
• torch==1.7.0
• ogb==1.2.3
• dgl==0.5.3
• cython==0.29.21
• tensorboardX==2.1

Prepare

pip3 install -r requirement.txt
python3 cython_sampler/setup.py build_ext -i
mkdir log

How to run

python3 train.py --logdir <log_folder> --dataset <dataset> --sampler <sampler> --model <model> --sample_interval <DeltaT> --neighbor_limit <k> --gamma 0.4 --eta 0.01 --epochs 200 --hidden1 256 --batchsize 256 --learning_rate 0.001 --dropout 0.1 

• <log_folder>: The folder to save TB file and running data under the root folder './log'.
• dataset: CoraFull / ogbn-arxiv / ogbn-products
• sampler: BanditSampler / Thanos
• model: GCN / GAT
• sample_interval: $$ \Delta_{T} $$, the interval for reinitialization of sampler. Set -1 for BanditSampler to turn off the reinitialization. Setting 0 will reinitialize sampler every epoch.
• neighbor_limit: k, the number of neighbor to be sampled.
• gamma: $$ \gamma $$
• eta: $$ \eta $$
• hidden1: the dimension of hidden embedding
• epochs: the number of epoches

Command Example for BanditSampler

python3 train.py --logdir log_corafull --dataset CoraFull --sampler BanditSampler  --model GCN --logger_name 1 --sample_interval -1 --neighbor_limit 3 --gamma 0.4 --etas 0.01 --epochs 300 --hidden1 256 --batchsize 256 --learning_rate 0.001 --dropout 0.1 --noadd_selfloop 

python3 train.py --logdir log_corafull --dataset CoraFull --sampler BanditSampler  --model GAT --logger_name 1 --sample_interval -1 --neighbor_limit 3 --gamma 0.4 --etas 0.01 --epochs 300 --hidden1 256 --batchsize 256 --learning_rate 0.001 --dropout 0.1 --add_selfloop 

Command Example for Thanos and Thanos.M

python3 train.py --logdir log_corafull --dataset CoraFull --sampler Thanos  --model GCN --logger_name 1 --sample_interval 2000 --neighbor_limit 3 --gamma 0.2 --etas 0.1 --epochs 300 --hidden1 256 --batchsize 256 --learning_rate 0.001 --dropout 0.1 --noadd_selfloop 

python3 train.py --logdir log_corafull --dataset CoraFull --sampler Thanos  --model GAT --logger_name 1 --sample_interval 2000 --neighbor_limit 3 --gamma 0.2 --etas 0.1 --epochs 300 --hidden1 256 --batchsize 256 --learning_rate 0.001 --dropout 0.1 --add_selfloop 

Check Results

tensorboard --logdir log/<log_folder>

Reference

If our paper and code are useful for your research, please cite the following article:

@article{zhang2021biased,
  title={A biased graph neural network sampler with near-optimal regret},
  author={Zhang, Qingru and Wipf, David and Gan, Quan and Song, Le},
  journal={arXiv preprint arXiv:2103.01089},
  year={2021}
}