Why neg_samples can take the neighbor nodes but not only the disparate nodes about graphsage HOT 2 CLOSED

Hi,

Yes you are right in that the negative sampler could possibly sample nearby nodes.
However, in the context of large graphs, as is the main focus of the paper, it is highly unlikely that the samples are neighbors of the node which we compute representations. In practice we do not find this to be a big issue. Of course if in some cases (e.g. when there are nodes with extremely large degree), you might want to modify the sampler such that neighboring nodes (or even nodes 2 hops away) are not sampled.

from graphsage.

In my large graph data, I also found that it will take a long long long time to random sample the neg_samples if need to make sure they are not the neighbors. So that, it's a nice choice to random sample nodes as the neg_samples without making sure they are not the neighbors since they are unlikely to be neighbors.
I will close this issue.
Thanks.

from graphsage.