Return latent edge features (when learned) #10539
Description
🚀 The feature, motivation and pitch
In GNN layers where latent edge features (or graph features) are learned, they should also be returned.
Currently, most implemented architectures only return the latent node features.
Maybe this could be done with a boolean return_latent_edges (similar to GATv2s return_attention_weights)?
Or is it not favorable anymore to learn new edge features based on the learned edge features of the previous GNN layers? This is how some books propose to deal with edge features, like the Bishop DL book.
Alternatives
No response
Additional context
No response