We are usually more conservative than what is required by NEP 29:

https://numpy.org/neps/nep-0029-deprecation_policy.html#support-table

Our current dependencies minimum supported versions are defined in:

https://github.com/scikit-learn/scikit-learn/blob/1.0.1/sklearn/_min_dependencies.py

I think it's fine to be conservative as long as CI and backport maintenance costs remain reasonable but we should specify what we want for the next release.

Let's consolidate our needs in the list below (feel free to update):

• bump to the Python 3.7 dep to 3.8 so as to only run the CI on Python 3.8 to 3.10
  • note: google colab is still using Python 3.7, I sent a comment on the relevant issue to warn their maintainers of our plans.
• bump numpy at least to 1.15 to be able to easily fix Use a stable sorting algorithm when selecting the max_features in TfidfVectorizer #21446

We can anticipate scikit-learn 1.1 to be released early-2022.

Are there other needs? Other things that could be greatly simplified by up-ing our minimal deps requirements?