Question: Is GCNConv only available for undirected graphs?

[Struggle-Forever]

GCNConv()

[rusty1s]

Yes, it should only be used for undirected graphs. If you want to apply it to directed graphs, I suggest you apply two GCNConv layers and sum up their outputs:

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

[Struggle-Forever]

Can I understand it like this.t like this： conv1(x, edge_index) is an undirected operation for graph encoders. conv2(x, edge_index.flip([0])) do directed information propagation via edge_index.flip([0])? What puzzles me is that the theoretical derivation of GCN is based on undirected graphs, is this feasible?

[rusty1s]

Yes, this is true. As you correctly mention, you will throw the theoretical derivation of GCN that way - my assumption is that it still works quite well. Alternatively, you can replace your GCNConv with SAGEConv, or have a look at this paper.

[Struggle-Forever]

thanks i finally adopt GAT, and the GraphSAGE is a good choice.