Closed eroell closed 1 month ago
There are some tests failing, solving this together with maybe some more overhauling - opened draft PR with examples to raise some suggestions, especially with examples, of what our ideas could materialize to.
Slightly bugged that the bootstrapping mean is typically larger than the single-computed distance, not 100% sure about reason and validity right now
Yeah, smaller test datasets are also fine. We have issues with some distances failing on smaller test datasets so we had to sadly keep them a bit on the larger side.
I'm fixing the scgen test now. The 3.11 pre-release fails due to something numba related.
Yeah, smaller test datasets are also fine. We have issues with some distances failing on smaller test datasets so we had to sadly keep them a bit on the larger side.
I'm fixing the scgen test now. The 3.11 pre-release fails due to something numba related.
(3 Failures caused by these changes right now I believe, to be fixed.)
PR Checklist
docs
is updatedDescription of changes Addresses issue #497. This PR introduces convenience functionality for using (random sampling with replacement) bootstrapping towards error estimation of distances between perturbations.
For the user, no breaking changes are introduced. The changes visible are
bootstrap: bool = False
,n_bootstrap: int = 100
,bootstrap_random_state: int = 0
in the methodsDistance.one_sided
andDistance.pairwise
Distance.bootstrap
, a bootsrapping analogon toDistance.__call__
Technical details
Additional context: Examples
__call__
vsbootstrap
onesided
pairwise