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

• 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.

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.

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.

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/.

This repository benefits a lot from HGB and GAMLP.