I’m trying to run RGCN for the link prediction task. However, I noticed that the code includes a propagation through test graph. I believe we shouldn’t have a propagation through the test graph as we might forget whatever we learned at training. It’s also some sort of cheating for link prediction because we are looking at all the nodes present in the graph and later we are going to ask about them.
What are your thoughts on this?
You are right! There is a work around to handle this in DGL. You can see more in this documentation: 6.2 Training GNN for Edge Classification with Neighborhood Sampling — DGL 0.6.1 documentation
1 Like