Use of target_names in example datasets is inconsistent

[mitar]

Description

It seems target_names sometimes correspond to values of categorical target and sometimes correspond to target colum name itself.

Steps/Code to Reproduce

from sklearn.datasets import load_linnerud
from sklearn.datasets import load_iris

print(load_linnerud()['target_names'])
print(load_iris()['target_names'])

Expected Results

I would expect consistency. Or that target_names target column names (even in the case of one column), or that it returns categorical target values (and then it should not exist for regression).

Actual Results

['Weight', 'Waist', 'Pulse']
array(['setosa', 'versicolor', 'virginica'], dtype='<U10')

Also note that once it is a list and once it is an array.

Versions

System:
    python: 3.6.7 (default, Oct 22 2018, 11:32:17)  [GCC 8.2.0]
executable: /usr/bin/python3
   machine: Linux-4.18.0-20-generic-x86_64-with-Ubuntu-18.04-bionic

BLAS:
    macros: HAVE_CBLAS=None
  lib_dirs: /usr/lib/x86_64-linux-gnu
cblas_libs: openblas, openblas

Python deps:
       pip: 19.0.3
setuptools: 41.0.1
   sklearn: 0.20.3
     numpy: 1.15.4
     scipy: 1.2.1
    Cython: None
    pandas: 0.23.4

[jnothman]

I'd be in favour of renaming the iris one to 'classes' on the basis that this is what we use elsewhere to map between class index and name.

[ogrisel]

I am afraid that this problem does not impact only iris. Several other classification dataset loader have the same issue: for instance fetch_lfw_people, fetch_lfw_pairs, fetch_20newsgroups, ...

The classification_report utility has a similar problem.

We need to conduct an audit of the list of things to update with git grep target_names + manual inspection.

For the name of the new attribute, I would be in favor of the more explicit class_names.

[thomasjpfan]

I noticed that fetch_openml uses target_names as a "column name" and not a list of unique classes.

I agree the first step is to move to a more explicit name such as class_names or follow our convention of classes (used in classifiers). In this case we would need to deprecate the current behavior of target_names.