FIX always scale continuous features to unit variance in mutual info #24747
Conversation
| Non-regression test for: | ||
| https://github.com/scikit-learn/scikit-learn/issues/23720 | ||
| """ | ||
| rng = np.random.RandomState(0) |
There was a problem hiding this comment.
I am wondering if this is not overkilled here.
There was a problem hiding this comment.
¯_(ツ)_/¯
Part of my question is to figure out what other people's feeling is for when to use the global seed thing and when not :)
There was a problem hiding this comment.
IMO, this is to test some numerical stability issues triggered in corner cases.
Here, I would not have necessarily done it because the two input vectors used to compute the MI will be used exactly with the same functions underneath.
betatim
left a comment
There was a problem hiding this comment.
LGTM, just one comment about global seeds or not
Co-authored-by: Tim Head <[email protected]>
|
Is it worth having a test that checks that the MI is symmetric also when the inputs are "correlated"? Something like: In [18]: d_c = c.astype(int)
In [21]: mutual_info_classif(
...: c[:, None], d_c, discrete_features=[False], random_state=123
...: )
Out[21]: array([0.92509398])
In [22]: mutual_info_regression(
...: d_c[:, None], c, discrete_features=[True], random_state=123
...: )
Out[22]: array([0.92509398]) |
Yep we can add this test as well. |
|
Maybe with a "high" and "low" MI test there is no need to test different seeds? |
OK, I see what you want to test. Let's try the fixture then. Indeed, it will use all random seeds only when requested (I always forget about it). |
Co-authored-by: Thomas J. Fan <[email protected]>
closes #23720
The continuous features in
Xshould be scaled to a unit variance independently ifyis continuous or discrete. This is an implementation detail reported in [1].[1] A. Kraskov, H. Stogbauer and P. Grassberger, “Estimating mutual information”. Phys. Rev. E 69, 2004.