GNN-LF/HF

WWW2021 : Interpreting and Unifying Graph Neural Networks with An Optimization Framework. [Best Paper Awards Nomination!]

Resources

[Paper] & [PPT] & [Video(in Chinese)] & [Video(in English)]

Environment Settings

• python == 3.8.5
• torch == 1.7.1
• numpy == 1.19.4
• scipy == 1.6.0
• networkx == 2.5
• scikit-learn == 0.24.0
• pandas == 1.2.0

Data

• Cora/Citeseer/Pubmed: Semi-Supervised Classifcation with Graph Convolutional Networks.
• ACM: Heterogeneous Graph Attention Network.
• Wiki-CS: Wiki-CS: A Wikipedia-Based Benchmark for Graph Neural Networks.
• MS-Academic: Predict then Propagate: Graph Neural Networks meet Personalized PageRank.

For wiki-cs dataset, please copy data files Here to path "/data/wiki/"

Usage

Input Parameters

• (required) -d/--dataset: name for datasets, i.e., cora/citeseer/pubmed/acm/wiki/ms.
• (required) -t: model type, PPNP = 0; GNN-LF = 1; GNN-HF = 2.
• (required) -f: propagation form, closed-form = 0; iterative-form = 1.
• -l/-labelrate: training rate, i.e., 20 nodes per class for training, default = 20.
• --niter: times for iteration, default = 10.
• --device: GPU number.
• --reg_lambda: weight for regularization, default = 5e-3.
• --lr: learning rate, default = 0.01

Command

• Closed-form GNN-LF:

python main.py --dataset=cora -t=1 -f=0 --device=0

• Iter-form GNN-LF:

python main.py --dataset=cora -t=1 -f=1 --device=0

Cite

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

@inproceedings{zhu2021interpreting,
  title={Interpreting and Unifying Graph Neural Networks with An Optimization Framework},
  author={Zhu, Meiqi and Wang, Xiao and Shi, Chuan and Ji, Houye and Cui, Peng},
  booktitle={Proceedings of the Web Conference 2021},
  pages={1215--1226},
  year={2021}
}

Contact

If you have any questions, please feel free to contact me with zhumeiqi@bupt.edu.cn