GroupBy(axis=1) Does Not Offer Implicit Selection By Columns Name(s)

[lebigot]

Code Sample, a copy-pastable example if possible

import pandas as pd
import numpy as np

df = pd.DataFrame(np.arange(12).reshape(3, 4), index=[0, 1, 0], columns=[10, 20, 10, 20])
df.index.name = "y"
df.columns.name = "x"

print df

print
print "Grouped along index:"
print df.groupby(by="y").sum()

print
print "Grouped along columns:"
# The following raises a KeyError even though  "x" is a column name
# (like "y" above, which is an index name):
df.groupby(by="x", axis=1).sum()

Problem description

The exception at the end is surprising: the intent is clearly to group by columns, on the "x" column label.

Furthermore, the documentation for groupby() seems to confirm this, as it states for the "by" argument that "A str or list of strs may be passed to group by the columns in self".

Expected Output

A dataframe with index [0, 1, 0] but grouped (and summed) columns [10, 20].

I wasn't able to test with the latest Pandas version, sorry!

Output of pd.show_versions()

INSTALLED VERSIONS

commit: None
python: 2.7.13.final.0
python-bits: 64
OS: Linux
OS-release: 2.6.32-642.15.1.el6.x86_64
machine: x86_64
processor: x86_64
byteorder: little
LC_ALL: None
LANG: en_US.UTF-8
LOCALE: None.None

pandas: 0.21.1
pytest: 3.2.3
pip: 9.0.1
setuptools: 28.8.0
Cython: 0.27.3
numpy: 1.13.3
scipy: 0.19.1
pyarrow: None
xarray: 0.8.2
IPython: 5.1.0
sphinx: 1.4.4
patsy: 0.4.1
dateutil: 2.6.1
pytz: 2017.2
blosc: None
bottleneck: None
tables: 3.4.4
numexpr: 2.6.5
feather: None
matplotlib: 2.1.2
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
lxml: None
bs4: 4.6.0
html5lib: 0.9999999
sqlalchemy: 1.2.18
pymysql: None
psycopg2: None
jinja2: 2.8
s3fs: None
fastparquet: None
pandas_gbq: None
pandas_datareader: None

[WillAyd]

What do you mean by crash? This gives a KeyError for x which makes sense because there is no x column.

Note we also don't support Python2 anymore

[topper-123]

The example that @lebigot shows, should return the same as df.T.groupby(by="x").sum().T and should return

>>> df.T.groupby(by="x").sum().T  # same as df.groupby(by="x", axis=1).sum()
x  10  20
y
0   2   4
1  10  12
0  18  20

Seems like a bug in groupby to me.

@lebigot if you'd want to investigate, that'd be great.

[WillAyd]

Hmm I think this example is somewhat confusing since x is not an actually value as much as a name. It's the column name when not transposed and the index name when transposed.

I guess by="x" allows for implicit selection by index name whereas we don't offer that same implicit selection by column name.

[topper-123]

I've never understood by to only be usable for index names, and because groupby allows axis=1, it would make sense to me that by follows the supplied axis, and not necessarily only follows the index.

So I see this as bug, given that axis=1 should give the same functionality as axis=0, but onlyshould work over the other axis?

[WillAyd]

Right there's some special casing going on for index names which I doubt is extended to column names. Might be somewhere around here:

pandas/pandas/core/groupby/grouper.py

Line 570 in 6e28b67

all_in_columns_index = all(

I think could logically extend for axis=1 to look at column names as well (in italics to disambiguate from labels)