ywjawmw/qw_loss

A Quasi-Wasserstein Loss for Learning Graph Neural Networks (QW loss)

QW-loss

• A Quasi-Wasserstein Loss for Learning Graph Neural Networks (WWW 2024) [https://arxiv.org/abs/2310.11762].

Training & Evaluation

• Bregman Divergence-based Solver.

python training.py  --lr 0.01 --weight_decay 0.0 --dataset Chameleon --net GAT --Original_ot ot --lambda_ 1.0 --q_linear_lr 0.002 --q_linear_delay 0.0005 

• Bregman ADMM-based Solver

python training_badmm.py  --lr 0.002 --weight_decay 0.0 --dataset Photo --net GIN --Original_ot ot --lambda_ 100.0 --q_linear_lr 0.001 --q_linear_delay 0.0005 --iter_num_theta 5 --iter_num_q 5

• Traditional GNN

python training.py  --lr 0.01 --weight_decay 0.0 --dataset Chameleon --net GAT --Original_ot Original

• The main GNN framework is based on BernNet[https://github.com/ivam-he/BernNet] and ChebNetII[https://github.com/ivam-he/ChebNetII].

Parameters

Original_ot indicates whether to choose QW-loss improved GNNs (Original_ot='ot') or traditional GNNs (Original_ot='Original').

lambda_ corresponds to the wight of Bergman Divergence.

q_linear_lr corresponds to the learning rate for label transport matrix F.

q_linear_delay corresponds to the weight decay for label transport matrix F.

Citation

If our work can help you, please cite it

@article{cheng2023quasi,
  title={A Quasi-Wasserstein Loss for Learning Graph Neural Networks},
  author={Cheng, Minjie and Xu, Hongteng},
  journal={arXiv preprint arXiv:2310.11762},
  year={2023}
}