Ordered GNN

The official implementation of the paper "Ordered GNN: Ordering Message Passing to Deal with Heterophily and Over-smoothing" (ICLR 2023).

Dependencies

This is the list of the package versions required for our experiments.

python==3.8.12
torch==1.11.0
torch_geometric==2.0.4
torch_sparse==0.6.13
torch_scatter==2.0.9
torch_cluster==1.6.0
torch_spline_conv==1.2.1
wandb==0.12.16

Run

We manage our experiments with wandb, to reproduce the results we reported in our paper, please follow these steps:

• Set up the environment variables. Below 2 environment variables $YOUR_WANDB_ENTITY$ and $YOUR_WANDB_PROJECT$ are your wandb username and the name of the project.

  export WANDB_entity=$YOUR_WANDB_ENTITY$
  export WANDB_project=$YOUR_WANDB_PROJECT$

• Choose best hyper-parameters you want to run with, and create wandb sweep with that file.

  We record the best hyper-parameters in folder best_params, you can index corresponding file by name. For example, over-smoothing_Cora_OrderedGNN_4.yaml stores hyper-parameters for experiment on over-smoothing problem on Cora dataset, with a 4 layers Ordered GNN model.

  python sweep.py --sweep_file=best_params/over-smoothing_Cora_OrderedGNN_4.yaml

• You will get an sweep ID $SWEEP_ID$ and sweep URL $SWEEP_URL$ from last step, like:

  Create sweep with ID: $SWEEP_ID$
  Sweep URL: $SWEEP_URL$

  then run below command will start runs with GPU. Parameter $INDEX_GPU$:$PARALLEL_RUNS$ indicate we will run $PARALLEL_RUNS$ runs in parallel with GPU $INDEX_GPU$.

  python agents.py --sweep_id=$SWEEP_ID$ --gpu_allocate=$INDEX_GPU$:$PARALLEL_RUNS$

• You can check the results in $SWEEP_URL$, a website hosted on wandb.ai.

Citation

If you found the provided code with our paper useful in your work, we kindly request that you cite our work.

@inproceedings{song2023ordered,
    title={Ordered GNN: Ordering Message Passing to Deal with Heterophily and Over-smoothing},
    author={Yunchong Song and Chenghu Zhou and Xinbing Wang and Zhouhan Lin},
    booktitle={The Eleventh International Conference on Learning Representations},
    year={2023}
}