Can GCNConv handle multi graphs?

[Laknath1996]

Does GCNConv or ChebConv operations support multi-dimensional edge weights? If not, can someone kindly let me know whether there are Graph Conv operators that support multiple edge attributes? The idea is to (if possible) use a GCNConv to explore features of a multi-graph (a graph where a given two nodes can have multiple undirected edges).

[wsad1]

• GCNConv and ChebConv don't support multi-dimensional edge attributes. But you could modify the message function in GCNConv to support that .

def message(self, x_j, edge_attr):
    # perform some operation like cat, so message depends on node and edge attributes.
    x_j=torch.cat([x_j, edge_attr] 
    return self.linear(x_j) # linear transform to out_channels. 

• Yes, there are layers like NNConv, PNAConv, GINEConv.. that support multiple edge attributes.

[rusty1s]

To follow-up on @wsad1, you have two more options:

1. You can simply represent your multi-graph by duplicating edges in edge_index.
2. You can iteratively compute the GNN output for each individual edge weight, e.g.:

out = 0
for i in range(num_edge_weights):
    out = out + self.convs[i](x, edge_index, edge_weight[:, i])

[Laknath1996]

Thank you very much @rusty1s and @wsad1! I'll try out the options you have mentioned above