[Roadmap] Better support for hypergraphs

[wsad1]

🚀 The feature, motivation and pitch

Following this discussion. The goal is to improve support for hypergraphs in pyg via:

• Add HyperGraphData object that inherits from BaseData and implements __inc__, __cat_dim__, num_nodes, subgraph etc for hyper graphs. The hyperedge_index will be of shape [2, num_hyper_edges*num_nodes_per_hyper_edge], see here for an example.[Hypergraph data object #7611]
• Test that DataLoader works with HyperGraphData. [Hypergraph data object #7611]
• Add a hyper graph dataset to Datasets
• Add hyper graph gnns like HGNN,
• Add an example that uses all of the above new modules.

[CodeTal]

@wsad1 Hello sir, I am a new open-source contributor and PyG user. I happened to see this issue and found that the hypergraph is a really interesting notion. Could I work on this issue?

[rusty1s]

I don't think @wsad1 is currently working on this. Just let me know when you would be available to work on this so that we avoid duplicated efforts.

[CodeTal]

@rusty1s OK. I will try to do something these days.

[CodeTal]

@rusty1s Thank you for letting me know!

[CodeTal]

Hello, I am trying to implement the HyperGraphData class first. I have gone through some relevant codes and formulated some initial ideas. I think most functions can be inherited from the Data class, and here are something that we need to change:

1. Re-name some parameters. For example, edge_index to hyperedge_index, adj to incidence
2. Re-implement some functions.

• Prior issues (Question about mini-batches training for hypergraph neural network #7365）mentioned that the __inc__ function should be refactored. Based on my understanding of how the tensors are collated during batching, the __inc__ of HyperGraphData should still return the num_nodes for the first row of hyperedge_index but return the num_hyperedges for the second row of hyperedge_index.
• The __cat_dim__ function could remain the same as the concatenation method doesn't change.
• The subgraph function should be changed because the expression of edge_index has changed.
• Other functions could remain the same as the original Data class.

I am new to this project, so I still have many confusions. If I am wrong, could you please provide some feedback when you are free? Thanks a lot!

[wsad1]

@CodeTal Thanks for showing interest in this. Yes you are right On all fronts. Are you planning to have HyperGraphData inherit from Data, I think that should be ok to start with.
Are you interested in sending in a PR with HyperGraphData and updates to __inc__, and subgraph?

[CodeTal]

Sure! I am working on this. Thank you for your reply!

[CodeTal]

Sorry for the delay, I was stuck by several engineering problems related to the project. I have created a pull request #7611 which contains a HyperGraphData with __inc__ and subgraph reimplemented. I found that there is no clear definition for a hypergraph's subgraph. So I chose a way to implement it (Please refer to the PR comment).

[wsad1]

Thanks @CodeTal will take a look by end of week.

[mehular0ra]

@wsad1 Hello, I am a new open-source contributor and a PyG user. Can I add a hyper graph dataset? I work on Neuroimaging data and can contribute ABIDE2 data. I can also contribute standard datasets as well.

[wsad1]

@mehular0ra sorry for the extremely delayed reply. We'd be happy to add your dataset to pyg.

[wsad1]

@mehular0ra will open another issue to track the remaining items.