self-loop as a feature for Graph Classification

[Eyemissing]

In my case, I have a scenario, where a self-loop is an important information used as a feature for my graph classification. As far as I understand this is in GNN not possible. Is this correct? what might be a work-around?
BR Mike

[rusty1s]

Most GNN layers will already implicitly assume a self-loop by adding a simple skip-connection. If you want to treat self-loops explicitly, you could do something like

out = conv1(x, edge_index[:, edge_index[0] != edge_index[1]]) + conv2(x, edge_index[:, edge_index[0] == edge_index[1]])