[python] Naming returned columns in Pandas aggregate function?

I'm having trouble with Pandas' groupby functionality. I've read the documentation, but I can't see to figure out how to apply aggregate functions to multiple columns and have custom names for those columns.

This comes very close, but the data structure returned has nested column headings:

data.groupby("Country").agg(
        {"column1": {"foo": sum()}, "column2": {"mean": np.mean, "std": np.std}})

(ie. I want to take the mean and std of column2, but return those columns as "mean" and "std")

What am I missing?

I agree with the OP that it seems more natural and consistent to name and define the output columns in the same place (e.g. as is done with tidyverse's summarize in R), but a work-around in pandas for now is to create the new columns with desired names via assign before doing the aggregation:

data.assign(
    f=data['column1'],
    mean=data['column2'],
    std=data['column2']
).groupby('Country').agg(dict(f=sum, mean=np.mean, std=np.std)).reset_index()

(Using reset_index turns 'Country', 'f', 'mean', and 'std' all into regular columns with a separate integer index.)

With the inspiration of @Joel Ostblom

For those who already have a workable dictionary for merely aggregation, you can use/modify the following code for the newer version aggregation, separating aggregation and renaming part. Please be aware of the nested dictionary if there are more than 1 item.

def agg_translate_agg_rename(input_agg_dict):
    agg_dict = {}
    rename_dict = {}
    for k, v in input_agg_dict.items():
        if len(v) == 1:
            agg_dict[k] = list(v.values())[0]
            rename_dict[k] = list(v.keys())[0]
        else:
            updated_index = 1
            for nested_dict_k, nested_dict_v in v.items():
                modified_key = k + "_" + str(updated_index)
                agg_dict[modified_key] = nested_dict_v
                rename_dict[modified_key] = nested_dict_k
                updated_index += 1
    return agg_dict, rename_dict

one_dict = {"column1": {"foo": 'sum'}, "column2": {"mean": 'mean', "std": 'std'}}
agg, rename = agg_translator_aa(one_dict)

We get

agg = {'column1': 'sum', 'column2_1': 'mean', 'column2_2': 'std'}
rename = {'column1': 'foo', 'column2_1': 'mean', 'column2_2': 'std'}

Please let me know if there is a smarter way to do it. Thanks.

If you want to have a behavior similar to JMP, creating column titles that keep all info from the multi index you can use:

newidx = []
for (n1,n2) in df.columns.ravel():
    newidx.append("%s-%s" % (n1,n2))
df.columns=newidx

It will change your dataframe from:

    I                       V
    mean        std         first
V
4200.0  25.499536   31.557133   4200.0
4300.0  25.605662   31.678046   4300.0
4400.0  26.679005   32.919996   4400.0
4500.0  26.786458   32.811633   4500.0

to

    I-mean      I-std       V-first
V
4200.0  25.499536   31.557133   4200.0
4300.0  25.605662   31.678046   4300.0
4400.0  26.679005   32.919996   4400.0
4500.0  26.786458   32.811633   4500.0

such as this kind of dataframe, there are two levels of thecolumn name:

 shop_id  item_id   date_block_num item_cnt_day       
                                  target              
0   0       30          1            31               

we can use this code:

df.columns = [col[0] if col[-1]=='' else col[-1] for col in df.columns.values]

result is:

 shop_id  item_id   date_block_num target              
0   0       30          1            31 

For pandas >= 0.25

The functionality to name returned aggregate columns has been reintroduced in the master branch and is targeted for pandas 0.25. The new syntax is .agg(new_col_name=('col_name', 'agg_func'). Detailed example from the PR linked above:

In [2]: df = pd.DataFrame({'kind': ['cat', 'dog', 'cat', 'dog'],
   ...:                    'height': [9.1, 6.0, 9.5, 34.0],
   ...:                    'weight': [7.9, 7.5, 9.9, 198.0]})
   ...:

In [3]: df
Out[3]:
  kind  height  weight
0  cat     9.1     7.9
1  dog     6.0     7.5
2  cat     9.5     9.9
3  dog    34.0   198.0

In [4]: df.groupby('kind').agg(min_height=('height', 'min'), 
                               max_weight=('weight', 'max'))
Out[4]:
      min_height  max_weight
kind
cat          9.1         9.9
dog          6.0       198.0

It will also be possible to use multiple lambda expressions with this syntax and the two-step rename syntax I suggested earlier (below) as per this PR. Again, copying from the example in the PR:

In [2]: df = pd.DataFrame({"A": ['a', 'a'], 'B': [1, 2], 'C': [3, 4]})

In [3]: df.groupby("A").agg({'B': [lambda x: 0, lambda x: 1]})
Out[3]:
         B
  <lambda> <lambda 1>
A
a        0          1

and then .rename(), or in one go:

In [4]: df.groupby("A").agg(b=('B', lambda x: 0), c=('B', lambda x: 1))
Out[4]:
   b  c
A
a  0  0

For pandas < 0.25

The currently accepted answer by unutbu describes are great way of doing this in pandas versions <= 0.20. However, as of pandas 0.20, using this method raises a warning indicating that the syntax will not be available in future versions of pandas.

Series:

FutureWarning: using a dict on a Series for aggregation is deprecated and will be removed in a future version

DataFrames:

FutureWarning: using a dict with renaming is deprecated and will be removed in a future version

According to the pandas 0.20 changelog, the recommended way of renaming columns while aggregating is as follows.

# Create a sample data frame
df = pd.DataFrame({'A': [1, 1, 1, 2, 2],
                   'B': range(5),
                   'C': range(5)})

# ==== SINGLE COLUMN (SERIES) ====
# Syntax soon to be deprecated
df.groupby('A').B.agg({'foo': 'count'})
# Recommended replacement syntax
df.groupby('A').B.agg(['count']).rename(columns={'count': 'foo'})

# ==== MULTI COLUMN ====
# Syntax soon to be deprecated
df.groupby('A').agg({'B': {'foo': 'sum'}, 'C': {'bar': 'min'}})
# Recommended replacement syntax
df.groupby('A').agg({'B': 'sum', 'C': 'min'}).rename(columns={'B': 'foo', 'C': 'bar'})
# As the recommended syntax is more verbose, parentheses can
# be used to introduce line breaks and increase readability
(df.groupby('A')
    .agg({'B': 'sum', 'C': 'min'})
    .rename(columns={'B': 'foo', 'C': 'bar'})
)

Please see the 0.20 changelog for additional details.

Update 2017-01-03 in response to @JunkMechanic's comment.

With the old style dictionary syntax, it was possible to pass multiple lambda functions to .agg, since these would be renamed with the key in the passed dictionary:

>>> df.groupby('A').agg({'B': {'min': lambda x: x.min(), 'max': lambda x: x.max()}})

    B    
  max min
A        
1   2   0
2   4   3

Multiple functions can also be passed to a single column as a list:

>>> df.groupby('A').agg({'B': [np.min, np.max]})

     B     
  amin amax
A          
1    0    2
2    3    4

However, this does not work with lambda functions, since they are anonymous and all return <lambda>, which causes a name collision:

>>> df.groupby('A').agg({'B': [lambda x: x.min(), lambda x: x.max]})
SpecificationError: Function names must be unique, found multiple named <lambda>

To avoid the SpecificationError, named functions can be defined a priori instead of using lambda. Suitable function names also avoid calling .rename on the data frame afterwards. These functions can be passed with the same list syntax as above:

>>> def my_min(x):
>>>     return x.min()

>>> def my_max(x):
>>>     return x.max()

>>> df.groupby('A').agg({'B': [my_min, my_max]})

       B       
  my_min my_max
A              
1      0      2
2      3      4