yule-BUAA / HGConv

codes of HGConv model for Heterogeneous Graph Learning

Hybrid Micro/Macro Level Convolution for Heterogeneous Graph Learning

This experiment is based on Stanford OGB (1.2.4) benchmark. The description of "Hybrid Micro/Macro Level Convolution for Heterogeneous Graph Learning" is available here. The steps are:

1. run python preprocess_ogbn_mag.py to preprocess the original ogbn_mag dataset. As the OGB-MAG dataset only has input features for paper nodes, for all the other types of nodes (author, affiliation, field), we use the metapath2vec model to generate their structural features.

2. run python train_ogbn_mag.py to train the model.

3. run python eval_ogbn_mag.py to evaluate the model.

Environments:

• PyTorch 1.7.0
• DGL 0.5.2
• PyTorch Geometric 1.6.1
• OGB 1.2.4
• tqdm
• numpy

Selected hyperparameters:

  num_heads              INT      Number of attention heads                          8
  hidden_units           INT      Dimension of hidden units for each head            32
  n_layers               INT      Number of GNN layers                               2
  learning_rate          FLOAT    Learning rate                                      0.001
  dropout                FLOAT    Dropout rate                                       0.5
  residual               BOOL     Whether to use the residual connection             True

Hyperparameters could be found in the args variable in train_ogbn_mag.py file and you can adjust them when training the model. When evaluating the model, please make sure the args in eval_ogbn_mag.py keep the same to those in the training process.

Reference performance for the OGB-MAG dataset:

Model	Test Accuracy	Valid Accuracy	# Parameter	Hardware
HGConv	0.5045 ± 0.0017	0.5300 ± 0.0018	2,850,405	NVIDIA TITAN Xp (12GB)

Citation

Please consider citing our paper when using the code.

@article{yu2020hybrid,
  title={Hybrid Micro/Macro Level Convolution for Heterogeneous Graph Learning},
  author={Yu, Le and Sun, Leilei and Du, Bowen and Liu, Chuanren and Lv, Weifeng and Xiong, Hui},
  journal={arXiv preprint arXiv:2012.14722},
  year={2020}
}