Towards Deeper Graph Neural Networks

This repository is an official PyTorch implementation of DAGNN in "Towards Deeper Graph Neural Networks" (KDD2020). Our implementation is mainly based on PyTorch Geometric, a geometric deep learning extension library for PyTorch.

For more insights, (empirical and theoretical) analysis, and discussions about deeper graph neural networks, please refer to our paper.

Meng Liu, Hongyang Gao, and Shuiwang Ji. Towards Deeper Graph Neural Networks.

Other unofficial implementations:

• An implementation in DGL [PyTorch]
• An implementation in GraphGallery [PyTorch]

Reference

@inproceedings{liu2020towards,
  title={Towards Deeper Graph Neural Networks},
  author={Liu, Meng and Gao, Hongyang and Ji, Shuiwang},
  booktitle={Proceedings of the 26th ACM SIGKDD International Conference on Knowledge Discovery \& Data Mining},
  year={2020},
  organization={ACM}
}

Requirements

• PyTorch
• PyTorch Geometric >= 1.3.1
• NetworkX
• tqdm

Note that the versions of PyTorch and PyTorch Geometric should be compatible and PyTorch Geometric is related to other packages, which need to be installed in advance. It would be easy by following the installation instruction.

PyTorch Geometric 1.3.1 was used in this code. If you have a newer version installed already, you may encounter an error about "GCNConv.norm" when running this code. Refer to this issue for a possible solution. (2020.8.12 update: This issue has been solved in the current code. Now, our code works for PyTorch Geometric >= 1.3.1.)

Run

• To evaluate DAGNN on OGB dataset ogbn-arxiv, run

CUDA_VISIBLE_DEVICES=0 python main_ogbnarxiv.py

Dataset	#Para	Validation	Test
ogbn-arxiv	43857	72.90±0.11	72.09±0.25

• To reproduce our results in Table 2 and 3, run

bash run.sh