[BUG] `GroupVariableImportanceMixin` should not use `sorted(self.features_groups.keys())`

[jpaillard]

A bug was introduced in the class: GroupVariableImportanceMixin.
The private attribute _features_groups_ids is generated by iterating through: sorted(self.features_groups.keys()). This creates a discrepancy between the order of groups in features_groups (which is the order expected by the user) and _features_groups_ids , which is used to create the $X^{-j}, X^j$ sets and then the importance array. This is illustrated below; the user expects 'dummy group' to be the second group, but its importance will actually be located at the first index of the importance array.

import numpy as np
import pandas as pd
from sklearn.linear_model import LinearRegression

from hidimstat import CFI

X = pd.DataFrame(np.random.rand(100, 3), columns=["a", "b", "c"])
y = np.random.randn(100)

features_groups = {
    "group 1": ["a", "b"],
    "dummy group": ["c"],
}

estimator = LinearRegression()
estimator.fit(X, y)
cfi = CFI(estimator=estimator, features_groups=features_groups)
cfi.fit(X, y)
print(cfi.features_groups)
print(cfi._features_groups_ids) 

out:

>>> features_groups: {'group 1': ['a', 'b'], 'dummy group': ['c']}
>>> _features_groups_ids [[2], [0, 1]]

[bthirion]

The problem is that .keys() would return a random order.
You mean that we should have a list next to the dictionary to keep the initial ordering ?

[jpaillard]

I see.
Yes, we need something like that or an attribute feature_group_names that allows us to put a name on each importance value.