About Planetoid edge_index utilization

Hi!

Thank you again for the flexible and interesting implementation of SEAL and the possibility of using custom datasets! I really enjoy it!

In the seal implementation, you are using the Planetoid dataset. May I ask why you are replacing the edge_index by split_edge['train']['edge'].t(), cf. edge index. Is this line specific to the Planetoid dataset?

[bits-glitch]

Thank you very much for the explanation! Now, I have finally understood why you do not apply data.edge_index = split_edge['train']['edge'].t() to the ogb datasets. I discovered that the edge_index in the ogb_datasets is already reduced (i.e. it already only contains training edges, and not all edges of the complete graph), which I was not aware of.

For example, if we take ogbl-ddi, we get:

• Train Split Edge Shape torch.Size([1067911, 2])
• Test Split Edge Shape torch.Size([133489, 2])
• Valid Split Shape torch.Size([133489, 2])

If I print the data object after data = dataset[0], I get the following result: Data(edge_index=[2, 2135822]) --> 2135822 / 2 = 1067911 (directed links), thus containing only the training links.

Thanks again for the explanation!

[muhanzhang]

Because the OGB datasets already have pre-specified train/test/val splits, and you can use dataset.get_edge_split() to retrieve it from dataset. In contrast, you need to do the splitting for custom datasets yourself using do_edge_split(). The OGB datasets have a data = dataset[0] which contains the observed network. Uinsg dataset.get_edge_split() you can also get its train/val/test edge splits. The train edges are subset of the observed network.