Directed GraphSAGE code sample

[pmoriano]

Hello,
Thanks for your hard work and making PyG a first class library for Geometric DL.

I wonder if you can point out to a sample of code implementing and using graphSAGE for directed graphs. I am unsure on how to extend the current GraphSAGE layer to support in- and out-neighbors.

Thanks for the help.

[rusty1s]

One way to utilize any GNN operator for directed graphs is to utilize two GNN layers for each edge direction. This way, each direction will not share any weights with the reverse one:

x = conv1(x, edge_index).relu() + conv2(x, edge_index.flip([0])).relu()

Let me know if that helps you!

[pmoriano]

@rusty1s thanks for the quick reply. It helps a lot.
I also found that the issues #397 and #1845 help to understand the implications of using convolutional layers when dealing with directed graphs.
I hope this maybe of guidance to others.