Measuring improvements against using features only

[mtlive]

Is there anyway to find out how much the graph data has contributed to prediction results? I mean how do you know the performance gains are from using the graph structure and not because of CNN model? Is there anyway to try GraphSAGE model or something close to it on features only?

[RexYing]

0 layers of message passing would be feature-only baseline right? E.g. for node classification, you simply need to learn an MLP that takes in node features and outputs the class.

[mtlive]

I tried this approach: I removed all edges and added self loopes for all nodes. Is it correct?

[RexYing]

That would work as well!