On Series with CategoricalIndex, __getitem__ not equal to .loc

[kernc]

Code Sample, a copy-pastable example if possible

>>> s = pd.Series([2, 1, 0], index=pd.CategoricalIndex([2, 1, 0]))
>>> s.index
CategoricalIndex([2, 1, 0], categories=[0, 1, 2], ordered=False, dtype='category')
>>> s[[0, 1, 2]].equals(s.loc[[0, 1, 2]])  # Should be True
False
>>> s[[0, 1, 2]].equals(s.iloc[[0, 1, 2]])  # Should be False
True

Problem description

For Series with CategoricalIndex, __getitem__ indexing behaves differently than .loc. On other-indexed series, these accessors return the same result:

>>> s = pd.Series([2, 1, 0], index=[2, 1, 0])
>>> s[[0, 1, 2]].equals(s.loc[[0, 1, 2]])  # Should be True
True
>>> s[[0, 1, 2]].equals(s.iloc[[0, 1, 2]])  # Should be True
False

Expected Output

>>> s = pd.Series([2, 1, 0], index=pd.CategoricalIndex([2, 1, 0]))
>>> s[[0, 1, 2]].equals(s.loc[[0, 1, 2]])
True

Output of pd.show_versions()

0.19.0+479.git

[jorisvandenbossche]

You have encountered a specific gotcha of getitem indexing. When you have a numerical index, it is label-based indexing (the second example you give). But for other types of indexes, it is positional indexing (eg try with strings).
The questions of course here is then: is a Categorical index with numerical categories a numerical index or a 'other type' of index?

If we say it is not a numerical index (which I think we should do, as the way of indexing of a CategoricalIndex should IMO not depend on the type of the categories), the current behaviour is correct. But I admit this is rather confusing ..

[jorisvandenbossche]

For some more discussion on getitem indexing, see #9595. The issue you raise here is yet another example of where the behaviour of getitem can be confusing/unpredictable due to the mixing of both label-based and positional indexing.

[jorisvandenbossche]

It seems that this issue only occurs for list-like indexers, as a scalar label works correctly. With the example of above:

In [53]: s[0]
Out[53]: 0    <---- label based

In [54]: s[[0]]
Out[54]: 
2    2      <---- position based
dtype: int64

In [55]: s.loc[[0]]
Out[55]: 
0    0
dtype: int64