Why are the positive samples generated twice on LinkNeighborSampler?

[daviddavo]

Using LinkNeighborLoader on a Heterogeneous graph with binary negative sampling will return an edge store like this (batch size=8).

(user, vote, proposal)={
     edge_index=[2, 15],
     input_id=[8],
     edge_label_index=[2, 16],
     edge_label=[16],
   },

It makes sense that edge_label_index is twice the batch size, as it has 8 positive samples and 8 negative samples (which can be selected using edge_label). But then what is edge_index for? Why does it have 15 pairs? Why not 8? If I disable negative sampling, there would be exactly 8 samples (but edge_label_index still exists).

Are the "positive samples" generated twice?

[akihironitta]

I remember I encountered a similar issue and I think #7791 fixed it. Would you mind trying it again with the master branch or nightly release?