How to use ImbalancedSampler for edge classfication and node classfication

[lfangyu09]

Hi,

Thank you so much for providing this wonderful package. I have recently been working on Edge classification and Node classification. It is binary classification but the dataset is quite imbalanced. So I would like to use ImbalancedSampler to preprocess the train dataset. But I faced two issues:

(1) Edge classification.
How to use ImbalancedSampler to process the train dataset for edge classification? I did these

train_sampler = pyg.loader.ImbalancedSampler(train_dataset)
train_loader = pyg.loader.DataLoader(train_dataset, batch_size=1, shuffle=True, num_workers=2, sampler=train_sampler)

But it returns this error:

It looks like ImbalancedSampler is based on Graph classification. So, how to use ImbalancedSampler to process the train dataset for edge?

(2) Node classification.
How to use ImbalancedSampler to process the train dataset for node classification? I did these

sampler = ImbalancedSampler(data.y)
loader = NeighborLoader(data, input_nodes=data.train_mask,
                        batch_size=64, num_neighbors=[-1, -1],
                        sampler=sampler, ...)

This looks like for each data rather than the whole dataset. So, how to use ImbalancedSampler to process the train dataset for nodes?

Thanks a lot.

[rusty1s]

1. Currently, ImbalancedSampler expects node-level or graph-level labels. You can also use it for edge-level classification, in which case you have to input tensors though, i.e., ImbalancedSampler(data.edge_label).
2. This looks totally fine, right? NeighborLoader can only operate on single graphs only anyway.