NeighborLoader imbalanced sampling question

[EvanUp]

Hey, I have two loader questions:

1. If num_neighbors is [30] and some nodes in a batch have fewer than 30 neighbors, how does NeighborLoader maintain consistent batch sizes?

2. Do any loaders in PyG support oversampling nodes in cases where node classes are imbalanced?

[rusty1s]

Hi and sorry for the late reply.

1. NeighborLoader does not oversample neighbors. If there exists fewer than 30 neighbors for a particular node, its full neighborhood will be sampled.
2. Similar to vanilla PyTorch, you should be able to pass in a custom sampler for oversampling/undersampling, see here. Let me know if that works for you.

[EvanUp]

Thank you! Really appreciate the response and the library! I hadn't realized NeighborLoader could able node indices and not just booleans - that made everything a lot easier! I should read the documentation more thoroughly next time :)