Description

check_is_fitted has false positives on custom subclasses with private attributes.

I believe check_is_fitted should not look at variables with a leading underscore because 1) that is Python's convention for private attributes, and 2) the scikit-learn API specifies fitted attributes as those with a trailing underscore (and there is no specification regarding leading underscores, correct?).

Backtracking from PR #14545 where the new check logic was introduced, I noticed that the check for leading underscore was added to cover these two modules:

• sklearn/neighbors/base.py
• sklearn/neighbors/lof.py

But perhaps these modules are actually not following the API specification?

Steps/Code to Reproduce

class MyPCA(PCA):
    def __init__(self, ...):  # omitted arguments for brevity
        super().__init__(...)
        self._my_private_attr = 42

mypca = MyPCA()
check_is_fitted(mypca)  # does not raise NotFittedError

Expected Results

check_is_fitted raises NotFittedError even on custom subclasses that have private attributes following the Python convention of a leading underscore.

Actual Results

NotFittedError is not raised.

Versions

>>> import sklearn; sklearn.show_versions()

System:
    python: 3.7.3 (default, Aug  8 2019, 19:40:58)  [GCC 5.4.0 20160609]
executable: /media/ale/data/education+research/code/baikal/venv/bin/python
   machine: Linux-4.4.0-170-generic-x86_64-with-debian-stretch-sid

Python dependencies:
       pip: 19.3.1
setuptools: 40.8.0
   sklearn: 0.22
     numpy: 1.17.4
     scipy: 1.3.3
    Cython: None
    pandas: None
matplotlib: None
    joblib: 0.14.0

Built with OpenMP: True