Subgraph dataset implementation

[dongkwan-kim]

Hi, I am implementing codes for subgraph datasets in Subgraph Neural Networks. I need some advices before creating a PR.

In InMemoryDataset, we first create data_list and call collate to get Data and slices. However, in subgraph prediction tasks, there are one supergraph (edges edge_index and node features x) and a set of subgraphs (node indices x_index). Edges in subgraphs can be induced from the supergraph using utils.subgraph.

Since edges in the global graph are not given to each subgraph instance, we cannot collate them into a single Data object. In my latest version, I made two separate member variables in InMemoryDataset: self.global_data for the supergraph, self.data for ​a set of subgraphs (w/ self.collate).

There does not seem to be an example of subgraph datasets in PyG (I might miss them), so I would like to discuss what the best form will be. Thanks in advance.

[rusty1s]

This depends on where you actually want to do the subgraph processing. I think you have two options:

1. Store subgraphs in separate data objects and collate them afterwards (good runtime, but bad memory consumption)
2. Store a single data object via self.collate, and override len and get in InMemoryDataset:

@property
def len(self):
    return x  # The number of subgraphs

def get(self, idx):
    # Perform subgraph computation here and return the sub data.

[dongkwan-kim]

I see. Thank you for the suggestion.
It would be nice if we let the users choose the processing method for their own use cases.
I will send the PR when the code is ready. Thanks again.