How can I implement a GNN which consider the edge weight in aggregation?

[xin-ran-w]

Hello, I have an idea to implement a GNN, which predicts node-level features (each node feature's dimension is 1).

Assuming that, there is a vertice $v$ (orange node in the figure) and its neighbours $\mathcal{N}(v)$ (blue nodes in the figure).
There are two relation types between $v$ and its neighbours.
The orange line whose value is $0.6$ represents the improve relation. It will increase the value of vertice $v$ through the message passing.
The blue line whose value is $-0.4$ represents the impress relation. It will decrease the value of vertice $v$ through the message passing.

I do not know which GNN I should use.

[rusty1s]

I suggest using GraphConv for this, which supports both positive and negative edge weights.

[xin-ran-w]

Thank you, I have tried GCN, however, I do not know how to apply edge weight in GCN.

[rusty1s]

conv = GraphConv(16, 32)
conv(x, edge_index, edge_weight)

where edge_weight is a vector holding the weight for each edge.

[xin-ran-w]

Thank you, I'll try! If it's effective, I will mark this as an answer.