Question about LinkNeighborLoader

[ruoyeruolan]

edge_label_index = train_data["user", "rates", "movie"].edge_label_index
edge_label = train_data["user", "rates", "movie"].edge_label
edge_label.unique()  #  tensor([0., 1.])

train_loader = LinkNeighborLoader(
        data=train_data,
        num_neighbors=[30] * 2,
        neg_sampling_ratio=2,
        edge_label_index=(("user", "rates", "movie"), edge_label_index),
        
        edge_label=edge_label,
        
        batch_size=128,
        shuffle=True )

sampled_data = next(iter(train_loader))
sampled_data["user", "rates", "movie"].edge_label  # tensor[1, 1, 0, 2...]

When I use the parameter edge_index, the edge_label inside train_loader will have an extra 2., but my edge_label is only 1, 0, and it works fine if I don't use it. Why?

python=3.12.4
torch=2.6.0
torch_geometric=2.6.1

[ruoyeruolan]

I already know the reason, because I have performed negative sampling in the split function, and negative sampling again in the loader function, which leads to this result. It is not recommended to perform double negative sampling like this.