Link prediction on multiple graph with custom negative edge sampling

[GianlucaDeStefano]

Hi all, I have recently been migrating to Pytorch Geometric from another library and am curious about how I should handle the following use case.

I have multiple separate graphs and want to train a link prediction model.
Each of these graphs is logically disconnected from the others. I want to avoid sampling negative edges that would form a bridge between any two of these graphs; as for my problem, it would be illogical and would actually damage the results.

In PyTorch geometric, what is the best way of managing these graphs for my task? I was looking at creating my custom dataset class and using diagonal batching. Is this the best way?

Furthermore, is there any prebuilt functionality to guarantee the sampling-of-negative-edges described above?

Thank you for any input you may give me,
Gianluca

[rusty1s]

I think you can use the normal DataLoader for this to batch your multiple graphs together, but just need to make sure to not sample negative links between two different graphs. There is a utils.batched_negative_sampling functionality which you can call inside your training loop to achieve this. Hope this helps!