Whether to support hypergraph neural networks, such as HGNN

[yaoysyao]

Whether to support hypergraph neural networks, such as HGNN(https://arxiv.org/abs/1809.09401)

[wsad1]

We have a hypergraphconv, do you think we can use a similar input representation to implement the HGNN?
It would be nice to support this. Happy to accept contributions.

[yaoysyao]

Thank you for your reply and thank you for your contribution to PYG，I think it is more difficult in hypergraph neural networks to construct hypergraphs to facilitate model training, if pyg has a related method of constructing hypergraphs, I think you can try to use pyg to construct hypergraphs to complete tasks such as classification, now hypergraph neural networks have been used by more and more researchers, how to simply construct hypergraph neural networks is a problem for many researchers

[wsad1]

Right. So to support hypergraphs better in pyg, we'll might need the following

1. HyperGraphData which will mostly inherit from Data and override some methods.
2. A HyperGraphLoader that can create minibatches of HyperGraphData from a list.
3. New HyperGraph GNNs.
4. Several example Datasets of HyperGraphData.
5. Maybe even a HyperGraphNeighborLoader.

These are just some initial thoughts open to suggestions.
cc: @rusty1s.

[yaoysyao]

Right. So to support hypergraphs better in pyg, we'll might need the following

1. HyperGraphData which will mostly inherit from Data and override some methods.
2. A HyperGraphLoader that can create minibatches of HyperGraphData from a list.
3. New HyperGraph GNNs.
4. Several example Datasets of HyperGraphData.
5. Maybe even a HyperGraphNeighborLoader.

These are just some initial thoughts open to suggestions. cc: @rusty1s.

Thanks for the reply, you are quite right，but I think not only supergraph GNNS, due to the current use of attention mechanism more and more frequently, may also need hypergraph attention neural network, and because more and more researchers begin to use graph neural networks to solve NLP-related problems, you can try to support NLP-related tasks

[yaoysyao]

In addition, I have another question I want to ask, whether the dataset in the PYG includes a dataset of text classification, I have checked the relevant source code, it seems that I have not found a dataset for text classification, can I provide a related dataset, or can I provide a relevant demo to facilitate me to create my own dataset to solve the text classification task

[rusty1s]

@wsad1 @yaoysyao I am definitely open for better hypergraph support, but IMO this is already well supported by DataLoader as long as you use a customized Data variant that increases hyperedge_index according to the shape of the underlying incidence matrix.

[yaoysyao]

Thank you for your reply, thank you for creating such a good tool, according to your reply, do you mean that as long as I can construct a graph structure that conforms to the PYG format, even if it is a hypergraph, I can use PYG for graph classification or node classification?

[rusty1s]

Yes, that is correct. You can use HyperGraphConv for all of these tasks. I am happy to include more hyperedge GNN layers in the future.