GCNConv GateGraphConv on heterogeneous graphs

[diegocervera]

Hi,

Is there a reason why GCNConv, GateGraphConv don't work in bipartite graphs? Is it just a matter of them not being implemented for bipartite graphs or is there an major underlying reason?

Thanks!

[wsad1]

GCNConv expects the source and destination nodes to be of the same type and have the same feature dimension, it cannot support bipartite graphs where source and destination nodes are not of the same type and might have different feature dimension.

Bottom line, its not a matter of not being implemented, the GCNConv operator was not built to work on bipartite graphs.

[muety]

I was asking myself the exact same question, so thanks @diegocervera @wsad1 for clarifying this. However, it's still not entirely clear to me, why GCNConv can't support different feature dimensions. Couldn't Θ simply map from the source node's feature dimension to the one of the destination node?

[rusty1s]

We could certainly use disjoint weight matrices for source and destination node features, but this wouldn't align anymore with the original definition of GCN.

[muety]

Thanks for the reply! There was a misunderstanding on my end. I was only thinking about message passing from node type B (source) to node type A (destination), where a Θ of size (d_B, d_A) would have worked. But of course, messages can be passed in reverse as well along the same type of edges, in which case we'd have to have a second set of weights of size (d_A, d_B) Sorry for bothering!