Why is validation/test set needed for unsupervised setting? about graphsage HOT 2 OPEN

Thanks for this question. I am also in the same situation (assuming you mean 'val' and 'test' attributes).
I also intend to use the embeddings for link prediction; I can see no reason why they shouldn't be useful. I think that the task of node classification is quite tightly coupled in the code so simply using the unsupervised setting to get the embeddings is unexpectedly difficult.

from graphsage.

Yes I mean 'val' and 'test' attributes. I figured the crash can be fixed by simply setting 'val' attribute as 'True' for some random nodes.. I guess validation against a loss function in unsupervised setting also makes sense but when the goal is learn an embeddings for each every node in the network it is unclear the role of 'train' and 'val' nodes..

from graphsage.