DHYPR model in PyG nn library

🚀 The feature, motivation and pitch

DHYPR is a new model developed by @hongluzhou, and myself focused on directed graph representation learning and can be used for several tasks such as link prediction, link sign prediction, and node classification you can read the rest of the paper here: https://arxiv.org/pdf/2112.11734.pdf.

Implementation: implementing this project would involve writing a transform to create a k-order matrix such as the implementation I've written here: https://github.com/advil64/dhypr-lib/blob/main/dhypr/datasets/custom_transforms.py as well as functions to perform message passing in hyperbolic space (these are still a work in progress to integrate into PyTorch geometric). The functions to perform hyperbolic message passing will likely be similar to the euclidian message passing classes already implemented in PyG but will obviously have different mathematical properties involving converting vectors to the hyperbolic space.

As you can see from the paper, our method is very effective in learning embeddings for directed graphs, especially those with cycles, compared to previous methods such as GCN and HGCN. We hope you'll find this work useful enough to add to PyG so that more people can use our methodologies.

Alternatives

No response

Additional context

Work on this is still underway; this issue was created to gauge interest and see if it's something that the maintainers would be interested in adding to the PyG repo. I am trying my best to make sure that our functions work seamlessly with other features, such as data loading and mini batching in PyG; if you want to take a look at the initial model's GitHub you can find it here: https://github.com/hongluzhou/dhypr. If you have done something similar and would like to help, please email me at advith.chegu@rutgers.edu or find me on the PyG slack!