Excluding target training edges from computational graph in Edge Data Loader

Is it possible to exclude the target training edges from the computational graph when using an Edge Data Loader? At the moment it looks like the only way to exclude the target training edges is by specifying the reverse edges. But what if I don’t have reverse edges, but still want to remove the target training edges to avoid information leakeage?

Thanks

1 Like

This is answered in Slack and I’m putting the answer here for bookkeeping:

For now you could work around it by specifying the reverse IDs to be the same as the original edge ID itself:

reverse_id=torch.arange(g.num_edges())

I’m adding a self option in [Feature] Add exclude self option for EdgeDataLoader by BarclayII · Pull Request #3122 · dmlc/dgl · GitHub.

1 Like