pd.Series.groupby is said to have only optional parameters

[tknuth]

s = pd.Series(x)
m = s.groupby().count() # does not work
m = s.groupby(s.values).count() # works

In https://github.com/pydata/pandas/blob/master/pandas/core/generic.py, by=None should be corrected to just by as None always leads to an error.

   def groupby(self, by=None, axis=0, level=None, as_index=True, sort=True,
            group_keys=True, squeeze=False):

        from pandas.core.groupby import groupby
        axis = self._get_axis_number(axis)
        return groupby(self, by, axis=axis, level=level, as_index=as_index,
            sort=sort, group_keys=group_keys, squeeze=squeeze)

See this http://stackoverflow.com/questions/17929426/groupby-for-pandas-series-not-working

[jorisvandenbossche]

by=None is possible, eg if you provide level, you don't need to specify by.

But we could probably provide a better error message here. Something like "You have to at least specify one of 'by' or 'level' "

Current error: TypeError: 'NoneType' object is not callable

[jorisvandenbossche]

Thinking about this again, I am a little bit confused:

The signature is: DataFrame.groupby(by=None, axis=0, level=None, ..), so there is a default of axis=0, and the explanation of level states "If the axis is a MultiIndex (hierarchical), group by a particular level or levels".
So when you don't have a MultiIndex, but a single Index, I would assume that you don't need to specify the level kwarg. So in that reasoning, df.groupby(axis=0) should just group by the row index?

In [39]: df = pd.DataFrame(np.random.randn(5,5), index=[0,0,1,1,1])

In [40]: df.groupby(axis=0, level=0).sum()
Out[40]:
          0         1         2         3         4
0 -2.219825  0.787332 -0.863655  0.877530  3.664685
1  0.766315 -2.294101 -1.636917 -0.205543  1.535448

In [41]: df.groupby(axis=0).sum()
-> TypeError: 'NoneType' object is not callable

So, or this should actually groupby by the index, or the documentation should be updated.

@hayd @jreback

[hayd]

Hmmmm. To me this looks much less readable than (something like):

df.groupby(df.index).sum()

I don't think this being the default is sensible (I think it would lead to confusion/bugs as well as not usually being a regular use case) ... ?

Saying that, definitely the error message should be fixed! Perhaps say, groupby a column(s), df.index or ...just (perish the thought) see the groupby docs!

[jorisvandenbossche]

@hayd yep, I can agree with that, being more explicit is better.

But then, I would argue that axis should not have a default of 0, so that df.groupby(df.index) would be equivalent to df.groupby(axis=0) when having a single index.
The problem with that is that you then can't specify only the level in df.groupby(level=1) when having a MultiIndex, but always also will have to specify axis: df.groupby(axis=0, level=1), as axis is no longer 0 by default.

A use case is to remove duplicates in your index, but indeed, not the most standard thing to use groupby for.

[hayd]

I would argue that this minor inconsistency (requiring level=0 to be passed, though I'm not sure I even agree it's inconsistent!) is a good compromise rather than allowing .groupby().

drop_duplicates is a bad example (there's function for that), but I know I've done this before for something and it can be useful, it's just infrequently required (which I think is good reason not to use it as the "default" a. you rarely use it b. you'd have to look up what it's doing when you do see it).

There's reason to have axis=0 by default, as we are grouping row-wise (be it on index or a column), it doesn't mean groupby the index!

That is to say, I like the current behaviour (both are explicit in what they are doing):

df.groupby(level=0) == df.groupby(df.index)

and leaving groupby blank should definitely raise

df.groupby() # reader has no idea what this should do

[jorisvandenbossche]

@hayd Ah yes, forgot about that meaning of axis=0!

So let's just say then: the docstring of level should be clarified that it is not only for a MultiIndex + add an explanation of the axis kwarg (there is none now)

And to further nitpick a bit :-) drop_duplicates is not fully a bad examples, as the drop_duplicates method looks at duplicate rows (which I first always forget) and not index, there is also a drop_duplicates method for index, but that is a bit less straightforward to use (as you have to re-index with the new index) + you can only drop duplicates, while with groupby you can also say you want the mean of the duplicates for example.

[jreback]

axis is just a modifier of of the level argument. Doesn't have meaning unless that is specified.

So in reality, you have to specify:

by or level/axis

would not allow just a bare, axis=0 because then it ONLY works if the index is an Index, better to be explict and have the user pass the actual axis

[springcoil]

This all feels like turtles all the way down :)

[jorisvandenbossche]

@jreback I think @hayd is right, that axis has indeed another meaning as well: on wich axis is the grouping applied, as you can also combine axis with by (and not only with level):

In [61]: df = pd.DataFrame(np.random.randn(5,5))

In [62]: df.groupby(by=pd.Series([1,1,1,2,2]), axis=0).sum()
Out[62]:
          0         1         2         3         4
1  3.685961  0.176227  0.422497  0.688798  0.341932
2  0.397360 -1.918284 -0.498714  3.177563  0.379342

In [63]: df.groupby(by=pd.Series([1,1,1,2,2]), axis=1).sum()
Out[63]:
          1         2
0  1.749975 -0.624686
1 -0.862650  1.331078
2  3.397360  0.324337
3 -1.941983  1.222833
4 -0.077655  2.334072

[jreback]

@jorisvandenbossche that look right, so axis is a modifier to by or level (which are the required parms).