In the GAT paper there’s a definition about neighborhood nodes (Ni), quote “these will be exactly the first-order neighbors of i (including i)”. Based on this, it seems like the message passing should always involve self-loops. However, in the source code of GATConv, I notice that there does not exist a self-message-passing process. The edge attentions are about the nodes without the source node itself. Is there anything that i missed? Thanks in advance!
The common practice to include self-loop message passing is to add self-loop to the input graph beforehand. Check out dgl.add_self_loop — DGL 2.3 documentation
This topic was automatically closed 30 days after the last reply. New replies are no longer allowed.