GCN_Norm for directed graph

[qubit0]

WebKB dataset has a directed graph and also has a few self-loop. How do we use gcn_norm here? What is the difference between source_to_target and target_to_source ? While adding a self-loop, does gcn_norm check if there is already a self-loop or not (It seems like add_remaining_self_loop is already doing that)?

If I use a naive implementation of GCN_norm (for example this one ) then it will just add an extra self-loop (even when there is already a self-loop) and also uses target_to_sourcelike calculation.

[rusty1s]

In theory, gcn_norm is only well-defined on undirected graphs since it takes the degrees for both source and destination node into account. Ofc, it will still give you a reasonable output for directed graphs, but you may need to choose whether to do the computation based on the in-degree (target_to_source) or out-degree (source_to_target). gcn_norm will check if there is already a self-loop present and will not another one for these nodes.