Open ghost opened 9 years ago
http://arxiv.org/pdf/1108.1427.pdf
This paper might be of interest to you. It mentions the trade-off of computational time and finite sample efficiency.
I think you are right, it does add a certain amount of bias, but the computational time necessary to do the whole graph would be insane.
How would using the signal subgraph instead of the whole graph affect bias?