BLEND for Graph Classification

This repository is an implementation / slight modification of the Beltrami Flow and Neural Diffusion on Graphs (BLEND) model proposed by Chamberlain et al. (2021). The official source code mainly focuses on node classification, while this repository uses BLEND for graph classification on the ShapeNet dataset. Most of the code is directly copied from the official code.

Requirements

The code has been tested running under Python 3.7.13.

Main Modifications from Official BLEND

• models/data.py includes the ShapeNet dataset
• gen_pos_encodings.py generates positional encodings for ShapeNet and saves them in data/pos_encodings
• models/GNN.py changes from node classification to graph classification task

Experiments

Due to the large quantity of the ShapeNet dataset compared to the original datasets for node classification, I only got to run the model on a couple categories of ShapeNet so far.

For example, to test the model on the "Cap" and "Rocket" categories of ShapeNet, run the following command:

python gen_pos_encodings.py --shapenet_data Cap,Rocket

Performance

• Test accuracy of ~73.23% on the single category of "Cap"

Further Work

• Test the model on the entire ShapeNet dataset, rather than a few categories

Reference

Chamberlain, B.P., Rowbottom, J., Eynard, D., Giovanni, F.D., Dong, X., & Bronstein, M.M. (2021). Beltrami Flow and Neural Diffusion on Graphs. Proceedings of the Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS) 2021. https://doi.org/10.48550/arXiv.2110.09443

@article
{chamberlain2021blend,
  title={Beltrami Flow and Neural Diffusion on Graphs},
  author={Chamberlain, Benjamin Paul and Rowbottom, James and Eynard, Davide and Di Giovanni, Francesco and Dong Xiaowen and Bronstein, Michael M},
  journal={Proceedings of the Thirty-fifth Conference on Neural Information Processing Systems (NeurIPS) 2021, Virtual Event},
  year={2021}
}

License

The official repository is under Apache License 2.0.