Change MultiIndex repr ?

[jorisvandenbossche]

From #13443 (comment)

The current MultiIndex representation looks like this:

In [15]: mi = pd.MultiIndex.from_tuples([('A', 1), ('A', 2),
    ...:                                 ('B', 3), ('B', 4)])

In [16]: mi
Out[16]:
MultiIndex(levels=[[u'A', u'B'], [1, 2, 3, 4]],
           labels=[[0, 0, 1, 1], [0, 1, 2, 3]])

So this shows the underlying labels and levels. Personally, I don't find this a very good repr, because:

• The shown level and labels are actually (internal) concepts of MultiIndex a lot of user do not have to think about.
• It also does not give you the best idea how the MultiIndex looks like at a glance IMO (since you have to look at the values of the labels to know what the order of the levels will be).

So the question is: is there a better alternative?

We could show tuples:

MultiIndex([('A', 1), ('A', 2), ('B', 3), ('B', 4)])

Or the individual levels:

MultiIndex([['A', 'A', 'B', 'B'], [1, 2, 3, 4]])

or ..

Historical note: in older versions, there was a difference between the repr and str of a MultiIndex:

In [1]: mi = pd.MultiIndex.from_tuples([('A', 1), ('A', 2),
   ...:                                         ('B', 3), ('B', 4)])

In [2]: mi
Out[2]:
MultiIndex(levels=[[u'A', u'B'], [1, 2, 3, 4]],
           labels=[[0, 0, 1, 1], [0, 1, 2, 3]])

In [3]: print mi

A  1
   2
B  3
   4

In [6]: pd.__version__
Out[6]: '0.14.1'

but seems to been disappeared (on purpose or not).

[J535D165]

It would be nice to have the possibility to change the representation through pandas.options.

[J535D165]

Any thoughts on the desired behaviour @jorisvandenbossche? I think that tuples are a good choice.

The tuple output is also close to the workaround many people use:

In [15]: x = numpy.arange(10000)

In [16]: pandas.MultiIndex.from_arrays([x, x]).values
Out[16]: 
array([(0, 0), (1, 1), (2, 2), ..., (9997, 9997), (9998, 9998),
       (9999, 9999)], dtype=object)

@rockg @hoechenberger

[jorisvandenbossche]

I personally also think that tuples would be good representation.

The main 'problem' with it is that something like MultiIndex([('A', 1), ('A', 2), ('B', 1)]) looks like code that can be ran as well, while this is not a valid construction. Which might be confusing (and also goes against the python idiom of the repr being unambiguous (but that's something we are already violating with other reprs as well)).

[topper-123]

I've made a suggestion that works locally but would appreciate input as to the exact format.

I too like the tuple format the best, but the tuples should IMO be vertically stacked, so the user at a glance can see both the individual level (vertically) and each row (horizontally).

Various attributes should also be shown at the bottom, to mirror CategoricalIndex.

An example with a reasonably complex MultiIndex:

>>> n = 100
>>> c = pd.CategoricalIndex(list('abcdefghijklmnop' * n), name='a')
>>> i = pd.Index(c.codes, name='b')
>>> pd.MultiIndex.from_arrays([c, i])
MultiIndex([ ('a', 0),
             ('b', 1),
             ('c', 2),
             ('d', 3),
             ('e', 4),
             ('f', 5),
             ('g', 6),
             ('h', 7),
             ('i', 8),
             ('j', 9),
            ...
             ('g', 6),
             ('h', 7),
             ('i', 8),
             ('j', 9),
            ('k', 10),
            ('l', 11),
            ('m', 12),
            ('n', 13),
            ('o', 14),
            ('p', 15)],
           levels=[
             CategoricalIndex(['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l',
                  'm', 'n', 'o', 'p'],
                 categories=['a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', ...], ordered=False, name='a', dtype='category'),
             Int64Index([0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15], dtype='int64', name='b'),
           ],
           names=['a', 'b'],
           length=1600)

I like that levels has type information. The individual levels should concatenate their value on the first line, but otherwise I like this. Thoughts?