Basic pytorch geometric architecture for graph signal classification

[Laolga]

Hi! I'm very new to pytorch geometric, but my understanding would be that the most common applications are graph/node classifications while I'd be very interested in graph signal classification.
To clarify, an example could be a feature matrix x (n,m) where for each author (n authors) we have indicators if they participated in a paper (m papers). Then we know the citation network for all papers (m by m adjacency matrix or (2,m) edge list. And we want to predict departments for the authors (so y is (1,n)).
Of course, I could also reformulate the problem as a graph classification, but it does not look very memory efficient if every author uses exactly the same graph and this graph is very big (and possibly weighted).
I couldn't really find a similar setting in any of the examples. Please let me know if I missed something or there is an easy way to change one of the common examples to make it work in this setting.

[rusty1s]

To me, this looks like a node classification problem on a heterogeneous graph (you have a author<->paper graph and a paper<->paper graph). In fact, this is quite similar to the ogbn-mag dataset. You can find some examples on how to handle such a heterogeneous setting here.

[Laolga]

Thanks a lot! This is indeed very helpful.
Do you think it still makes sense to model the data as a heterogeneous graph if we assume that the abovementioned feature matrix is not space, but dense and very fat (m >> n)?

[rusty1s]

Personally, I would not interpret the mapping from authors to papers as a feature matrix, but instead as an additional edge connectivity, and rely on other reasonable node features (such as paper abstracts).

[Laolga]

Can this edge be weighted? I did not see so far a setting of R-GCN with weighted edges. Do you know any examples like that? Because the relation is not binary in my case, unfortunately, and not sparse either.
Thanks again for your help!

[rusty1s]

Edges can be weighted, and be processed by GNN ops such as GraphConv by passing in the edge_weight argument.