search

williamleif / GraphSAGE

Representation learning on large graphs using stochastic graph convolutions.
Other
3.42k stars 842 forks source link

Measuring improvements against using features only #94

Open mtlive opened 5 years ago

mtlive commented 5 years ago

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 commented 5 years ago

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 commented 5 years ago

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

RexYing commented 5 years ago

That would work as well!