Simple and Efficient Heterogeneous Graph Neural Network (SeHGNN)

The pre-print paper at: http://arxiv.org/abs/2207.02547

Requirements

1. Some mainstream neural network libraries for GNNs

• pytorch
• pytorch-geometric
• dgl

Please check your cuda version first and install the above libraries matching your cuda. If possible, we recommend to install the newest versions of these libraries.

2. Other dependencies

Install other requirements:

pip install -r requirements.txt

Compile and install sparse-tools. Under the folder ./sparse_tools/, run

python setup.py develop

sparse-tools is implemented for acceleration of label propagation for large dataset such as ogbn-mag.

Data preparation

For the preliminary experiments and experiments on four middle-scale datasets, please download datasets DBLP.zip, ACM.zip, IMDB.zip, Freebase.zip from the source of HGB benchmark, and extract content from these compresesed files under the folder './data/'.

For the experiments on the large dataset ogbn-mag, the dataset will be automatically downloaded from OGB Challenge.

Experiments

For the preliminary experiments on HAN and HGB in Section 4 of the paper, please refer to folders ./preliminary/HAN/ and ./preliminary/HGB/, respectively.

For the experiments on four middle-scale datasets in Section 6 of the paper, please refer to the folder ./middle/.

For the experiments on the large dataset ogbn-mag in Section 6 of the paper, please refer to the folder ./large/.

Acknowledgement

This repository benefits a lot from HGB and GAMLP.