We propose a novel framework PAS to automatically learn data-specific pooling architectures for graph classification task. Firstly, a unified framework consisting of four essential modules is designed. Based on this framework, an effective search space is designed by incorporating popular operations from existing human-designed architectures. To enable efficient architecture search, we develop a coarsening strategy to continuously relax the search space, thus a differentiable search method can be adopted.

Latest version of Pytorch-geometric(PyG) is required. More details can be found in here

torch-cluster==1.5.7  
torch-geometric==1.7.2
torch-scatter==2.0.6  
torch==1.6.0  
numpy==1.17.2  
hyperopt==0.2.5  
python==3.7.4

Step 1. Run the search process, given different random seeds. (The 2-layer GNN on DD dataset is used as an example)

python train_search.py  --data DD   --num_layers 2  --epochs 100

The results are saved in the directory exp_res, e.g., exp_res/DD.txt.

Step 2. Fine tune the searched architectures. You need specify the arch_filename with the resulting filename from Step 1.

python fine_tune.py --data DD --num_layers 2 --ft_weight_decay  --ft_dropout  --ft_mode 10fold --hyper_epoch 30 --epochs 100    --arch_filename  ./exp_res/DD.txt 

The searched architectures and hyper-parameters are provided: (DD dataset for example)

python reproduce.py --data DD --gpu 0

Please kindly cite our paper if you use this code:

@inproceedings{wei2021pooling,
  title={Pooling Architecture Search for Graph Classification},
  author={Wei, Lanning and Zhao, Huan and Yao, Quanming and He, Zhiqiang},
  booktitle={CIKM},
  year={2021}
}