Ping @glouppe @jmschrei

Looks good to me, +1

That should also work with bootstrapping, right? It does on my box.

Not sure if I understand https://github.com/scikit-learn/scikit-learn/pull/5240/files#r39152326. Why would it fail with changes?

Not sure if I understand https://github.com/scikit-learn/scikit-learn/pull/5240/files#r39152326. Why would it fail with changes?

For numerical reasons, the new test doesn't highlight some numerical discrepancy in the numerical importance. Our conclusion is that the current test are too weak to detect some numerical error.
@glouppe said that he would bring new tests to ensure correctness.

That should also work with bootstrapping, right? It does on my box.

This depends on the samples that are drawn for the 10 trees. I have preferred to be correct all the time instead of depending on some randomisations.

+1 for improving the stability of this particular test. I am merging.