yule-BUAA / LEGNN

codes of LEGNN for Semi-supervised Node Classification

Label-Enhanced Graph Neural Network for Semi-supervised Node Classification

The experiments are conducted on the ogbn-arxiv and ogbn-mag datasets on the Stanford OGB (1.3.2) benchmark. The description of "Label-Enhanced Graph Neural Network for Semi-supervised Node Classification" is available here.

Datasets:

We provide the preprocessed datasets at here, which should be put in the ./dataset folder.

To run the node classification task on ogbn-arxiv:

• run ./train_LEGNN/train_full_batch_LEGNN.py to get the results of LEGNN on ogbn-arxiv.
• run ./train_LEGNN_ASTrain/train_full_batch_LEGNN_ASTrain.py to get the results of LEGNN + AS-Train on ogbn-arxiv.

To run the node classification task on ogbn-mag:

1. in addition to downloading the preprocessed datasets, you could also run ./preprocess_data/preprocess_ogbn_mag.py to preprocess the original ogbn-mag dataset.

• run ./train_LEGNN/train_mini_batch_RGNN.py to get the results of LEGNN on ogbn-mag.
• run ./train_LEGNN_ASTrain/train_mini_batch_LEGNN_ASTrain.py to get the results of LEGNN + AS-Train on ogbn-mag.

Performance on ogbn-arxiv

Model	Test Accuracy	Valid Accuracy	# Parameter	Hardware
LEGNN	0.7337 ± 0.0007	0.7480 ± 0.0009	5,374,120	NVIDIA Tesla T4 (15 GB)
LEGNN + AS-Train	0.7371 ± 0.0011	0.7494 ± 0.0008	5,374,120	NVIDIA Tesla T4 (15 GB)

Performance on ogbn-mag

Model	Test Accuracy	Valid Accuracy	# Parameter	Hardware
LEGNN	0.5276 ± 0.0014	0.5443 ± 0.0009	5,147,997	NVIDIA Tesla T4 (15 GB)
LEGNN + AS-Train	0.5378 ± 0.0016	0.5528 ± 0.0013	5,147,997	NVIDIA Tesla T4 (15 GB)

Environments:

• PyTorch 1.8.1
• DGL 0.7.0
• PyTorch Geometric 2.0.1
• OGB 1.3.2
• tqdm
• numpy

Citation

Please consider citing our paper when using the codes.

@article{yu2022label,
  title={Label-Enhanced Graph Neural Network for Semi-supervised Node Classification},
  author={Yu, Le and Sun, Leilei and Du, Bowen and Zhu, Tongyu and Lv, Weifeng},
  journal={arXiv preprint arXiv:2205.15653},
  year={2022}
}