The problem

The current behavior of the LabelEncoder is to sort the variables when the mapping is performed. This happens because of the use of np.unique which returns a sorted array of unique values: See https://github.com/Elementa-Engineering/scikit-optimize/blob/master/skopt/space/transformers.py#L175-L177 and https://numpy.org/doc/stable/reference/generated/numpy.unique.html

For example:

>>> from skopt.space.space import Categorical

>>> c = Categorical(("c", "b", "a"), transform="label")
>>> c.transform(["a", "b", "c"])
[0, 1, 2]

Note that the returned labels are 0, 1 and 2 (equivalent to ("a", "b", "c") even if the specified order was ("c", "b", "a")). This can be counter-intuitive, especially when the order of the variable "means" something for the user.

Implemented Fix

This PR, implements a simple fix, which retains the order of the categorical dimensions. The expected behavior then becomes:

>>> from skopt.space.space import Categorical

>>> c = Categorical(("c", "b", "a"), transform="label")
>>> c.transform(["a", "b", "c"])
[2, 1, 0]

The order is conserved.

Same goes for numerical numbers:

from skopt.space.space import Categorical

c = Categorical((10, 30, 20), transform="label")
c.transform([10, 20, 30])
[0, 2, 1]