HubGT

This is the original implementation of HubGT: Scalable Graph Transformer with Decoupled Hierarchy Labeling.

Tech report is available at Appendix.pdf.

Data Preparation

Datasets are automatically downloaded by the code and stored in the data folder.

Environment setup

Install by Conda:

conda create -n gt python=3.12
conda activate gt

conda install pytorch=2.2 -c pytorch -c nvidia
conda install pyg=2.5.3 -c pyg -c conda-forge
conda install scipy optuna tqdm torchmetrics pandas scikit-learn cython=3.0
conda install -c conda-forge ogb

Precomputation

• Environment: CMake 3.16, C++ 14. Compile the Cython code for precomputation:

cd Precompute/ && python setup.py build_ext --inplace && cd ..

• Index is stored in the cache folder.

Run the experiments

• Run training and evaluation:

python main.py --data [dataset] --batch [batch_size] -ss [total size] -s0 [out label size] -s0g [global label size] -s1 [in label size]

For other configurations refer to: python main.py --help

• Use script scripts/tune.sh to tune hyperparameters, script scripts/run.sh to run the code with the best hyperparameters.

• Results are stored in the log folder, summary is stored in log/summary.csv.