MichailChatzianastasis / WLHN

Weisfeiler and Leman go Hyperbolic: Learning Distance Preserving Node Representations

Weisfeiler and Leman go Hyperbolic: Learning Distance Preserving Node Representations (WLHN)

This repository is the original implementation of the paper titled Weisfeiler and Leman go Hyperbolic: Learning Distance Preserving Node Representations, accepted at AISTATS2023.

Requirements

Code is written in Python 3.9 and requires:

• PyTorch 1.12
• PyTorch Geometric 2.1
• geoopt 0.5

Usage

To train our model WLHN, use the above command:

python tu_dataset.py 

Arguments:

--dataset "Dataset name"
--lr "Initial Learning rate"
--dropout "Dropout rate"
--batch-size "Input batch size for training"
--epochs "Number of epochs to train"
--hidden-dim "Size of hidden layer"
--tau "Tau value for DiffHypCon construction"
--depth "Depth of WL tree"
--classifier "Classifier (hyperbolic_mlr or logmap)
--hyperbolic-optimizer "Whether to use hyperbolic optimizer"

Cite

Please cite our paper if you use this code:

@inproceedings{nikolentzos2022weisfeiler,
  title={Weisfeiler and Leman go Hyperbolic: Learning Distance Preserving Node Representations},
  author={Nikolentzos, Giannis and Chatzianastasis, Michail and Vazirgiannis, Michalis},
  booktitle={Proceedings of the 26th International Conference on Artificial Intelligence and Statistics},
  year={2023}
}