This is the official source code repo of paper "Masked Graph Autoencoder with Non-discrete Bandwidths" in TheWebConf(WWW) 2024.

We explore a new paradigm of topological masked graph autoencoders with non-discrete masking strategies, named "bandwidths". We verify its effectiveness in learning network topology by both theory and experiment.

| 📄 Preprint version (full version) | 📖 Published version | 👁️‍🗨️ OpenReview | 💬 Blog |

See requirements.txt.

See run.ipynb for our experiment results. You can either run the model by this Jupyter file or by commands below in the terminal:

python train_link.py --dataset=<dataset_name> --use_cfg --device=<gpu_id>

<dataset_name>: Cora, Citeseer, Pubmed, Photo, Computers, CS, Physics

By --use_cfg, the best hyperparameters in the config/<dataset_name>.yml file are used by default.

python train_link_ogb.py --dataset=<dataset_name> --use_cfg --device=<gpu_id>

<dataset_name>: ogbl-collab, ogbl-ppa

python train_node.py --dataset=<dataset_name> --use_cfg --device=<gpu_id>

<dataset_name>: Cora, Citeseer, Pubmed, Photo, Computers, CS, Physics, Wiki-CS, ogbn-arxiv, ogbn-mag

Please cite our paper for your research if our paper helps:

@inproceedings{bandana,
  title={Masked Graph Autoencoder with Non-discrete Bandwidths}, 
  author={Ziwen, Zhao and Yuhua, Li and Yixiong, Zou and Jiliang, Tang and Ruixuan, Li},
  booktitle={Proceedings of the 33rd ACM Web Conference},
  year={2024},
  month={May},
  publisher={Association for Computing Machinery},
  address={Singapore, Singapore},
}

• MaskGAE