GraphSAGE for isolated nodes

[aure_bnp]

Hi everyone !

I am currently working with GraphSAGE module from dgl and I run some tests on GraphSAGE. Firstly, how does graphSAGE works when a node has no neighbours ? Since he cant aggregate the informations from his neighbours, he can’t compute its embedding ?

Secondly, I have tried to run graphSAGE on a small graph where there is only isolated nodes with differents features information about the nodes, and GraphSAGE does not return a fixed embedding for each isolated node. Is it normal or am i misunderstanding an elemnt of GraphSAGE algorithm ?

Thank you all !

[mufeili]

For the first question, if the module employs residual connection, then you can view that as an MLP.

The second question might also have something to do with the residual connection. In general, it does not make sense to apply a graph neural network to a graph without edges, i.e. a set of points.

[aure_bnp]

Thanks a lot for your fast response.
I understand better now, if I can view that as an MLP it explains why the output of the node embeddings are not the same on the graph without edges.

Thanks again !