search

williamleif / GraphSAGE

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

Train edge seperation for unsupervised training #79

Open AnuradhaSK opened 5 years ago

AnuradhaSK commented 5 years ago

When I was following the unsupervised training code, I found this train edge separation. Here, Train edges = All edges - (edges between a valid node to a valid node) - (edges between a valid node to a train node without a train neighbourhood ) Can you please clarify the reason for those edge set deduction?

RexYing commented 5 years ago

So in the training time, we also want to remove a subset of edges. This is not for test/validation. This is just so that the model does not know existence of these edges when trying to predict them. This could lead to better generalization.