论文:[Regularizing graph neural networks via consistency-diversity graph augmentations]<https://ojs.aaai.org/index.php/AAAI/article/view/20307/20066>
@inproceedings{bo2022regularizing,
title={Regularizing graph neural networks via consistency-diversity graph augmentations},
author={Bo, Deyu and Hu, BinBin and Wang, Xiao and Zhang, Zhiqiang and Shi, Chuan and Zhou, Jun},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
volume={36},
number={4},
pages={3913--3921},
year={2022}
}
以 Amazon Photo 数据集为例子
- 数据下载: 从 https://github.com/BUPT-GAMMA/NASA/tree/main/dataset 下载photo.npz文件,放到data_process/dataset/目录下
- 数据预处理与子图采样:
运行submit.sh进行数据预处理和spark采样,得到每条样本的子图
将得到的csv表
graph_features.csv,包含 'seed', 'graph_feature', 'node_id_list', 'label_list', 'flag_list' 三个字段,其中 'flag_list' 字段中以 'train', 'val', 'test' 标记训练集数据、验证集数据和测试集数据。然后,执行下述指令,转换flag_list的格式以便于后续处理:
sed -i 's/\btrain\b/0/g' graph_features.csv
sed -i 's/\tval/\t1/g' graph_features.csv
sed -i 's/test/2/g' graph_features.csv
python nasa.py
- 效果
- NASA 经调参,300 epoch, Amazon Photo, F1 ~ 0.92 左右。