Group By: split-apply-combine

By “group by” we are referring to a process involving one or more of the following steps

  • Splitting the data into groups based on some criteria
  • Applying a function to each group independently
  • Combining the results into a data structure

Of these, the split step is the most straightforward. In fact, in many situations you may wish to split the data set into groups and do something with those groups yourself. In the apply step, we might wish to one of the following:

  • Aggregation: computing a summary statistic (or statistics) about each group. Some examples:

    • Compute group sums or means
    • Compute group sizes / counts
  • Transformation: perform some group-specific computations and return a like-indexed. Some examples:

    • Standardizing data (zscore) within group
    • Filling NAs within groups with a value derived from each group
  • Filtration: discard some groups, according to a group-wise computation that evaluates True or False. Some examples:

    • Discarding data that belongs to groups with only a few members
    • Filtering out data based on the group sum or mean
  • Some combination of the above: GroupBy will examine the results of the apply step and try to return a sensibly combined result if it doesn’t fit into either of the above two categories

Since the set of object instance method on pandas data structures are generally rich and expressive, we often simply want to invoke, say, a DataFrame function on each group. The name GroupBy should be quite familiar to those who have used a SQL-based tool (or itertools), in which you can write code like:

SELECT Column1, Column2, mean(Column3), sum(Column4)
FROM SomeTable
GROUP BY Column1, Column2

We aim to make operations like this natural and easy to express using pandas. We’ll address each area of GroupBy functionality then provide some non-trivial examples / use cases.

See the cookbook for some advanced strategies

Splitting an object into groups

pandas objects can be split on any of their axes. The abstract definition of grouping is to provide a mapping of labels to group names. To create a GroupBy object (more on what the GroupBy object is later), you do the following:

>>> grouped = obj.groupby(key)
>>> grouped = obj.groupby(key, axis=1)
>>> grouped = obj.groupby([key1, key2])

The mapping can be specified many different ways:

  • A Python function, to be called on each of the axis labels
  • A list or NumPy array of the same length as the selected axis
  • A dict or Series, providing a label -> group name mapping
  • For DataFrame objects, a string indicating a column to be used to group. Of course df.groupby('A') is just syntactic sugar for df.groupby(df['A']), but it makes life simpler
  • A list of any of the above things

Collectively we refer to the grouping objects as the keys. For example, consider the following DataFrame:

In [1]: df = DataFrame({'A' : ['foo', 'bar', 'foo', 'bar',
   ...:                        'foo', 'bar', 'foo', 'foo'],
   ...:                 'B' : ['one', 'one', 'two', 'three',
   ...:                        'two', 'two', 'one', 'three'],
   ...:                 'C' : randn(8), 'D' : randn(8)})
   ...: 

In [2]: df
Out[2]: 
     A      B         C         D
0  foo    one  0.469112 -0.861849
1  bar    one -0.282863 -2.104569
2  foo    two -1.509059 -0.494929
3  bar  three -1.135632  1.071804
4  foo    two  1.212112  0.721555
5  bar    two -0.173215 -0.706771
6  foo    one  0.119209 -1.039575
7  foo  three -1.044236  0.271860

We could naturally group by either the A or B columns or both:

In [3]: grouped = df.groupby('A')

In [4]: grouped = df.groupby(['A', 'B'])

These will split the DataFrame on its index (rows). We could also split by the columns:

In [5]: def get_letter_type(letter):
   ...:     if letter.lower() in 'aeiou':
   ...:         return 'vowel'
   ...:     else:
   ...:         return 'consonant'
   ...: 

In [6]: grouped = df.groupby(get_letter_type, axis=1)

Starting with 0.8, pandas Index objects now supports duplicate values. If a non-unique index is used as the group key in a groupby operation, all values for the same index value will be considered to be in one group and thus the output of aggregation functions will only contain unique index values:

In [7]: lst = [1, 2, 3, 1, 2, 3]

In [8]: s = Series([1, 2, 3, 10, 20, 30], lst)

In [9]: grouped = s.groupby(level=0)

In [10]: grouped.first()
Out[10]: 
1    1
2    2
3    3
dtype: int64

In [11]: grouped.last()
Out[11]: 
1    10
2    20
3    30
dtype: int64

In [12]: grouped.sum()
Out[12]: 
1    11
2    22
3    33
dtype: int64

Note that no splitting occurs until it’s needed. Creating the GroupBy object only verifies that you’ve passed a valid mapping.

Note

Many kinds of complicated data manipulations can be expressed in terms of GroupBy operations (though can’t be guaranteed to be the most efficient). You can get quite creative with the label mapping functions.

GroupBy object attributes

The groups attribute is a dict whose keys are the computed unique groups and corresponding values being the axis labels belonging to each group. In the above example we have:

In [13]: df.groupby('A').groups
Out[13]: {'bar': [1L, 3L, 5L], 'foo': [0L, 2L, 4L, 6L, 7L]}

In [14]: df.groupby(get_letter_type, axis=1).groups
Out[14]: {'consonant': ['B', 'C', 'D'], 'vowel': ['A']}

Calling the standard Python len function on the GroupBy object just returns the length of the groups dict, so it is largely just a convenience:

In [15]: grouped = df.groupby(['A', 'B'])

In [16]: grouped.groups
Out[16]: 
{('bar', 'one'): [1L],
 ('bar', 'three'): [3L],
 ('bar', 'two'): [5L],
 ('foo', 'one'): [0L, 6L],
 ('foo', 'three'): [7L],
 ('foo', 'two'): [2L, 4L]}

In [17]: len(grouped)
Out[17]: 6

By default the group keys are sorted during the groupby operation. You may however pass sort=False for potential speedups:

In [18]: df2 = DataFrame({'X' : ['B', 'B', 'A', 'A'], 'Y' : [1, 2, 3, 4]})

In [19]: df2.groupby(['X'], sort=True).sum()
Out[19]: 
   Y
X   
A  7
B  3

In [20]: df2.groupby(['X'], sort=False).sum()
Out[20]: 
   Y
X   
B  3
A  7

GroupBy will tab complete column names (and other attributes)

In [21]: df
Out[21]: 
            gender     height      weight
2000-01-01    male  42.849980  157.500553
2000-01-02    male  49.607315  177.340407
2000-01-03    male  56.293531  171.524640
2000-01-04  female  48.421077  144.251986
2000-01-05    male  46.556882  152.526206
2000-01-06  female  68.448851  168.272968
2000-01-07    male  70.757698  136.431469
2000-01-08  female  58.909500  176.499753
2000-01-09  female  76.435631  174.094104
2000-01-10    male  45.306120  177.540920

In [22]: gb = df.groupby('gender')
In [23]: gb.<TAB>
gb.agg        gb.boxplot    gb.cummin     gb.describe   gb.filter     gb.get_group  gb.height     gb.last       gb.median     gb.ngroups    gb.plot       gb.rank       gb.std        gb.transform
gb.aggregate  gb.count      gb.cumprod    gb.dtype      gb.first      gb.groups     gb.hist       gb.max        gb.min        gb.nth        gb.prod       gb.resample   gb.sum        gb.var
gb.apply      gb.cummax     gb.cumsum     gb.fillna     gb.gender     gb.head       gb.indices    gb.mean       gb.name       gb.ohlc       gb.quantile   gb.size       gb.tail       gb.weight

GroupBy with MultiIndex

With hierarchically-indexed data, it’s quite natural to group by one of the levels of the hierarchy.

In [24]: s
Out[24]: 
first  second
bar    one      -0.575247
       two       0.254161
baz    one      -1.143704
       two       0.215897
foo    one       1.193555
       two      -0.077118
qux    one      -0.408530
       two      -0.862495
dtype: float64

In [25]: grouped = s.groupby(level=0)

In [26]: grouped.sum()
Out[26]: 
first
bar     -0.321085
baz     -0.927807
foo      1.116437
qux     -1.271025
dtype: float64

If the MultiIndex has names specified, these can be passed instead of the level number:

In [27]: s.groupby(level='second').sum()
Out[27]: 
second
one      -0.933926
two      -0.469555
dtype: float64

The aggregation functions such as sum will take the level parameter directly. Additionally, the resulting index will be named according to the chosen level:

In [28]: s.sum(level='second')
Out[28]: 
second
one      -0.933926
two      -0.469555
dtype: float64

Also as of v0.6, grouping with multiple levels is supported.

In [29]: s
Out[29]: 
first  second  third
bar    doo     one      1.346061
               two      1.511763
baz    bee     one      1.627081
               two     -0.990582
foo    bop     one     -0.441652
               two      1.211526
qux    bop     one      0.268520
               two      0.024580
dtype: float64

In [30]: s.groupby(level=['first','second']).sum()
Out[30]: 
first  second
bar    doo       2.857824
baz    bee       0.636499
foo    bop       0.769873
qux    bop       0.293100
dtype: float64

More on the sum function and aggregation later.

DataFrame column selection in GroupBy

Once you have created the GroupBy object from a DataFrame, for example, you might want to do something different for each of the columns. Thus, using [] similar to getting a column from a DataFrame, you can do:

In [31]: grouped = df.groupby(['A'])

In [32]: grouped_C = grouped['C']

In [33]: grouped_D = grouped['D']

This is mainly syntactic sugar for the alternative and much more verbose:

In [34]: df['C'].groupby(df['A'])
Out[34]: <pandas.core.groupby.SeriesGroupBy object at 0xa36a3a4c>

Additionally this method avoids recomputing the internal grouping information derived from the passed key.

Iterating through groups

With the GroupBy object in hand, iterating through the grouped data is very natural and functions similarly to itertools.groupby:

In [35]: grouped = df.groupby('A')

In [36]: for name, group in grouped:
   ....:        print(name)
   ....:        print(group)
   ....: 
bar
     A      B         C         D
1  bar    one -0.042379 -0.089329
3  bar  three -0.009920 -0.945867
5  bar    two  0.495767  1.956030
foo
     A      B         C         D
0  foo    one -0.919854 -1.131345
2  foo    two  1.247642  0.337863
4  foo    two  0.290213 -0.932132
6  foo    one  0.362949  0.017587
7  foo  three  1.548106 -0.016692

In the case of grouping by multiple keys, the group name will be a tuple:

In [37]: for name, group in df.groupby(['A', 'B']):
   ....:        print(name)
   ....:        print(group)
   ....: 
('bar', 'one')
     A    B         C         D
1  bar  one -0.042379 -0.089329
('bar', 'three')
     A      B        C         D
3  bar  three -0.00992 -0.945867
('bar', 'two')
     A    B         C        D
5  bar  two  0.495767  1.95603
('foo', 'one')
     A    B         C         D
0  foo  one -0.919854 -1.131345
6  foo  one  0.362949  0.017587
('foo', 'three')
     A      B         C         D
7  foo  three  1.548106 -0.016692
('foo', 'two')
     A    B         C         D
2  foo  two  1.247642  0.337863
4  foo  two  0.290213 -0.932132

It’s standard Python-fu but remember you can unpack the tuple in the for loop statement if you wish: for (k1, k2), group in grouped:.

Aggregation

Once the GroupBy object has been created, several methods are available to perform a computation on the grouped data.

An obvious one is aggregation via the aggregate or equivalently agg method:

In [38]: grouped = df.groupby('A')

In [39]: grouped.aggregate(np.sum)
Out[39]: 
            C         D
A                      
bar  0.443469  0.920834
foo  2.529056 -1.724719

In [40]: grouped = df.groupby(['A', 'B'])

In [41]: grouped.aggregate(np.sum)
Out[41]: 
                  C         D
A   B                        
bar one   -0.042379 -0.089329
    three -0.009920 -0.945867
    two    0.495767  1.956030
foo one   -0.556905 -1.113758
    three  1.548106 -0.016692
    two    1.537855 -0.594269

As you can see, the result of the aggregation will have the group names as the new index along the grouped axis. In the case of multiple keys, the result is a MultiIndex by default, though this can be changed by using the as_index option:

In [42]: grouped = df.groupby(['A', 'B'], as_index=False)

In [43]: grouped.aggregate(np.sum)
Out[43]: 
     A      B         C         D
0  bar    one -0.042379 -0.089329
1  bar  three -0.009920 -0.945867
2  bar    two  0.495767  1.956030
3  foo    one -0.556905 -1.113758
4  foo  three  1.548106 -0.016692
5  foo    two  1.537855 -0.594269

In [44]: df.groupby('A', as_index=False).sum()
Out[44]: 
     A         C         D
0  bar  0.443469  0.920834
1  foo  2.529056 -1.724719

Note that you could use the reset_index DataFrame function to achieve the same result as the column names are stored in the resulting MultiIndex:

In [45]: df.groupby(['A', 'B']).sum().reset_index()
Out[45]: 
     A      B         C         D
0  bar    one -0.042379 -0.089329
1  bar  three -0.009920 -0.945867
2  bar    two  0.495767  1.956030
3  foo    one -0.556905 -1.113758
4  foo  three  1.548106 -0.016692
5  foo    two  1.537855 -0.594269

Another simple aggregation example is to compute the size of each group. This is included in GroupBy as the size method. It returns a Series whose index are the group names and whose values are the sizes of each group.

In [46]: grouped.size()
Out[46]: 
A    B    
bar  one      1
     three    1
     two      1
foo  one      2
     three    1
     two      2
dtype: int64
In [47]: grouped.describe()
Out[47]: 
                C         D
0 count  1.000000  1.000000
  mean  -0.042379 -0.089329
  std         NaN       NaN
  min   -0.042379 -0.089329
  25%   -0.042379 -0.089329
  50%   -0.042379 -0.089329
  75%   -0.042379 -0.089329
...           ...       ...
5 mean   0.768928 -0.297134
  std    0.677005  0.898022
  min    0.290213 -0.932132
  25%    0.529570 -0.614633
  50%    0.768928 -0.297134
  75%    1.008285  0.020364
  max    1.247642  0.337863

[48 rows x 2 columns]

Note

Aggregation functions will not return the groups that you are aggregating over if they are named columns, when as_index=True, the default. The grouped columns will be the indices of the returned object.

Passing as_index=False will return the groups that you are aggregating over, if they are named columns.

Aggregating functions are ones that reduce the dimension of the returned objects, for example: mean, sum, size, count, std, var, describe, first, last, nth, min, max. This is what happens when you do for example DataFrame.sum() and get back a Series.

nth can act as a reducer or a filter, see here

Applying multiple functions at once

With grouped Series you can also pass a list or dict of functions to do aggregation with, outputting a DataFrame:

In [48]: grouped = df.groupby('A')

In [49]: grouped['C'].agg([np.sum, np.mean, np.std])
Out[49]: 
          sum      mean       std
A                                
bar  0.443469  0.147823  0.301765
foo  2.529056  0.505811  0.966450

If a dict is passed, the keys will be used to name the columns. Otherwise the function’s name (stored in the function object) will be used.

In [50]: grouped['D'].agg({'result1' : np.sum,
   ....:                   'result2' : np.mean})
   ....: 
Out[50]: 
      result2   result1
A                      
bar  0.306945  0.920834
foo -0.344944 -1.724719

On a grouped DataFrame, you can pass a list of functions to apply to each column, which produces an aggregated result with a hierarchical index:

In [51]: grouped.agg([np.sum, np.mean, np.std])
Out[51]: 
            C                             D                    
          sum      mean       std       sum      mean       std
A                                                              
bar  0.443469  0.147823  0.301765  0.920834  0.306945  1.490982
foo  2.529056  0.505811  0.966450 -1.724719 -0.344944  0.645875

Passing a dict of functions has different behavior by default, see the next section.

Applying different functions to DataFrame columns

By passing a dict to aggregate you can apply a different aggregation to the columns of a DataFrame:

In [52]: grouped.agg({'C' : np.sum,
   ....:              'D' : lambda x: np.std(x, ddof=1)})
   ....: 
Out[52]: 
            C         D
A                      
bar  0.443469  1.490982
foo  2.529056  0.645875

The function names can also be strings. In order for a string to be valid it must be either implemented on GroupBy or available via dispatching:

In [53]: grouped.agg({'C' : 'sum', 'D' : 'std'})
Out[53]: 
            C         D
A                      
bar  0.443469  1.490982
foo  2.529056  0.645875

Cython-optimized aggregation functions

Some common aggregations, currently only sum, mean, and std, have optimized Cython implementations:

In [54]: df.groupby('A').sum()
Out[54]: 
            C         D
A                      
bar  0.443469  0.920834
foo  2.529056 -1.724719

In [55]: df.groupby(['A', 'B']).mean()
Out[55]: 
                  C         D
A   B                        
bar one   -0.042379 -0.089329
    three -0.009920 -0.945867
    two    0.495767  1.956030
foo one   -0.278452 -0.556879
    three  1.548106 -0.016692
    two    0.768928 -0.297134

Of course sum and mean are implemented on pandas objects, so the above code would work even without the special versions via dispatching (see below).

Transformation

The transform method returns an object that is indexed the same (same size) as the one being grouped. Thus, the passed transform function should return a result that is the same size as the group chunk. For example, suppose we wished to standardize the data within each group:

In [56]: index = date_range('10/1/1999', periods=1100)

In [57]: ts = Series(np.random.normal(0.5, 2, 1100), index)

In [58]: ts = rolling_mean(ts, 100, 100).dropna()

In [59]: ts.head()
Out[59]: 
2000-01-08    0.779333
2000-01-09    0.778852
2000-01-10    0.786476
2000-01-11    0.782797
2000-01-12    0.798110
Freq: D, dtype: float64

In [60]: ts.tail()
Out[60]: 
2002-09-30    0.660294
2002-10-01    0.631095
2002-10-02    0.673601
2002-10-03    0.709213
2002-10-04    0.719369
Freq: D, dtype: float64

In [61]: key = lambda x: x.year

In [62]: zscore = lambda x: (x - x.mean()) / x.std()

In [63]: transformed = ts.groupby(key).transform(zscore)

We would expect the result to now have mean 0 and standard deviation 1 within each group, which we can easily check:

# Original Data
In [64]: grouped = ts.groupby(key)

In [65]: grouped.mean()
Out[65]: 
2000    0.442441
2001    0.526246
2002    0.459365
dtype: float64

In [66]: grouped.std()
Out[66]: 
2000    0.131752
2001    0.210945
2002    0.128753
dtype: float64

# Transformed Data
In [67]: grouped_trans = transformed.groupby(key)

In [68]: grouped_trans.mean()
Out[68]: 
2000   -7.561268e-17
2001   -4.194514e-16
2002   -1.362729e-16
dtype: float64

In [69]: grouped_trans.std()
Out[69]: 
2000    1
2001    1
2002    1
dtype: float64

We can also visually compare the original and transformed data sets.

In [70]: compare = DataFrame({'Original': ts, 'Transformed': transformed})

In [71]: compare.plot()
Out[71]: <matplotlib.axes.AxesSubplot at 0xa3511b2c>
_images/groupby_transform_plot.png

Another common data transform is to replace missing data with the group mean.

In [72]: data_df
Out[72]: 
            A         B         C
0    1.539708 -1.166480  0.533026
1    1.302092 -0.505754       NaN
2   -0.371983  1.104803 -0.651520
3   -1.309622  1.118697 -1.161657
4   -1.924296  0.396437  0.812436
5    0.815643  0.367816 -0.469478
6   -0.030651  1.376106 -0.645129
..        ...       ...       ...
993  0.012359  0.554602 -1.976159
994  0.042312 -1.628835  1.013822
995 -0.093110  0.683847 -0.774753
996 -0.185043  1.438572       NaN
997 -0.394469 -0.642343  0.011374
998 -1.174126  1.857148       NaN
999  0.234564  0.517098  0.393534

[1000 rows x 3 columns]

In [73]: countries = np.array(['US', 'UK', 'GR', 'JP'])

In [74]: key = countries[np.random.randint(0, 4, 1000)]

In [75]: grouped = data_df.groupby(key)

# Non-NA count in each group
In [76]: grouped.count()
Out[76]: 
      A    B    C
GR  209  217  189
JP  240  255  217
UK  216  231  193
US  239  250  217

In [77]: f = lambda x: x.fillna(x.mean())

In [78]: transformed = grouped.transform(f)

We can verify that the group means have not changed in the transformed data and that the transformed data contains no NAs.

In [79]: grouped_trans = transformed.groupby(key)

In [80]: grouped.mean() # original group means
Out[80]: 
           A         B         C
GR -0.098371 -0.015420  0.068053
JP  0.069025  0.023100 -0.077324
UK  0.034069 -0.052580 -0.116525
US  0.058664 -0.020399  0.028603

In [81]: grouped_trans.mean() # transformation did not change group means
Out[81]: 
           A         B         C
GR -0.098371 -0.015420  0.068053
JP  0.069025  0.023100 -0.077324
UK  0.034069 -0.052580 -0.116525
US  0.058664 -0.020399  0.028603

In [82]: grouped.count() # original has some missing data points
Out[82]: 
      A    B    C
GR  209  217  189
JP  240  255  217
UK  216  231  193
US  239  250  217

In [83]: grouped_trans.count() # counts after transformation
Out[83]: 
      A    B    C
GR  228  228  228
JP  267  267  267
UK  247  247  247
US  258  258  258

In [84]: grouped_trans.size() # Verify non-NA count equals group size
Out[84]: 
GR    228
JP    267
UK    247
US    258
dtype: int64

Note

Some functions when applied to a groupby object will automatically transform the input, returning an object of the same shape as the original. Passing as_index=False will not affect these transformation methods.

For example: fillna, ffill, bfill, shift.

In [85]: grouped.ffill()
Out[85]: 
            A         B         C
0    1.539708 -1.166480  0.533026
1    1.302092 -0.505754  0.533026
2   -0.371983  1.104803 -0.651520
3   -1.309622  1.118697 -1.161657
4   -1.924296  0.396437  0.812436
5    0.815643  0.367816 -0.469478
6   -0.030651  1.376106 -0.645129
..        ...       ...       ...
993  0.012359  0.554602 -1.976159
994  0.042312 -1.628835  1.013822
995 -0.093110  0.683847 -0.774753
996 -0.185043  1.438572 -0.774753
997 -0.394469 -0.642343  0.011374
998 -1.174126  1.857148 -0.774753
999  0.234564  0.517098  0.393534

[1000 rows x 3 columns]

Filtration

New in version 0.12.

The filter method returns a subset of the original object. Suppose we want to take only elements that belong to groups with a group sum greater than 2.

In [86]: sf = Series([1, 1, 2, 3, 3, 3])

In [87]: sf.groupby(sf).filter(lambda x: x.sum() > 2)
Out[87]: 
3    3
4    3
5    3
dtype: int64

The argument of filter must be a function that, applied to the group as a whole, returns True or False.

Another useful operation is filtering out elements that belong to groups with only a couple members.

In [88]: dff = DataFrame({'A': np.arange(8), 'B': list('aabbbbcc')})

In [89]: dff.groupby('B').filter(lambda x: len(x) > 2)
Out[89]: 
   A  B
2  2  b
3  3  b
4  4  b
5  5  b

Alternatively, instead of dropping the offending groups, we can return a like-indexed objects where the groups that do not pass the filter are filled with NaNs.

In [90]: dff.groupby('B').filter(lambda x: len(x) > 2, dropna=False)
Out[90]: 
    A    B
0 NaN  NaN
1 NaN  NaN
2   2    b
3   3    b
4   4    b
5   5    b
6 NaN  NaN
7 NaN  NaN

For dataframes with multiple columns, filters should explicitly specify a column as the filter criterion.

In [91]: dff['C'] = np.arange(8)

In [92]: dff.groupby('B').filter(lambda x: len(x['C']) > 2)
Out[92]: 
   A  B  C
2  2  b  2
3  3  b  3
4  4  b  4
5  5  b  5

Note

Some functions when applied to a groupby object will act as a filter on the input, returning a reduced shape of the original (and potentitally eliminating groups), but with the index unchanged. Passing as_index=False will not affect these transformation methods.

For example: head, tail.

In [93]: dff.groupby('B').head(2)
Out[93]: 
   A  B  C
0  0  a  0
1  1  a  1
2  2  b  2
3  3  b  3
6  6  c  6
7  7  c  7

Dispatching to instance methods

When doing an aggregation or transformation, you might just want to call an instance method on each data group. This is pretty easy to do by passing lambda functions:

In [94]: grouped = df.groupby('A')

In [95]: grouped.agg(lambda x: x.std())
Out[95]: 
      B         C         D
A                          
bar NaN  0.301765  1.490982
foo NaN  0.966450  0.645875

But, it’s rather verbose and can be untidy if you need to pass additional arguments. Using a bit of metaprogramming cleverness, GroupBy now has the ability to “dispatch” method calls to the groups:

In [96]: grouped.std()
Out[96]: 
            C         D
A                      
bar  0.301765  1.490982
foo  0.966450  0.645875

What is actually happening here is that a function wrapper is being generated. When invoked, it takes any passed arguments and invokes the function with any arguments on each group (in the above example, the std function). The results are then combined together much in the style of agg and transform (it actually uses apply to infer the gluing, documented next). This enables some operations to be carried out rather succinctly:

In [97]: tsdf = DataFrame(randn(1000, 3),
   ....:                  index=date_range('1/1/2000', periods=1000),
   ....:                  columns=['A', 'B', 'C'])
   ....: 

In [98]: tsdf.ix[::2] = np.nan

In [99]: grouped = tsdf.groupby(lambda x: x.year)

In [100]: grouped.fillna(method='pad')
Out[100]: 
                   A         B         C
2000-01-01       NaN       NaN       NaN
2000-01-02 -0.353501 -0.080957 -0.876864
2000-01-03 -0.353501 -0.080957 -0.876864
2000-01-04  0.050976  0.044273 -0.559849
2000-01-05  0.050976  0.044273 -0.559849
2000-01-06  0.030091  0.186460 -0.680149
2000-01-07  0.030091  0.186460 -0.680149
...              ...       ...       ...
2002-09-20  2.310215  0.157482 -0.064476
2002-09-21  2.310215  0.157482 -0.064476
2002-09-22  0.005011  0.053897 -1.026922
2002-09-23  0.005011  0.053897 -1.026922
2002-09-24 -0.456542 -1.849051  1.559856
2002-09-25 -0.456542 -1.849051  1.559856
2002-09-26  1.123162  0.354660  1.128135

[1000 rows x 3 columns]

In this example, we chopped the collection of time series into yearly chunks then independently called fillna on the groups.

Flexible apply

Some operations on the grouped data might not fit into either the aggregate or transform categories. Or, you may simply want GroupBy to infer how to combine the results. For these, use the apply function, which can be substituted for both aggregate and transform in many standard use cases. However, apply can handle some exceptional use cases, for example:

In [101]: df
Out[101]: 
     A      B         C         D
0  foo    one -0.919854 -1.131345
1  bar    one -0.042379 -0.089329
2  foo    two  1.247642  0.337863
3  bar  three -0.009920 -0.945867
4  foo    two  0.290213 -0.932132
5  bar    two  0.495767  1.956030
6  foo    one  0.362949  0.017587
7  foo  three  1.548106 -0.016692

In [102]: grouped = df.groupby('A')

# could also just call .describe()
In [103]: grouped['C'].apply(lambda x: x.describe())
Out[103]: 
A         
bar  count    3.000000
     mean     0.147823
     std      0.301765
     min     -0.042379
     25%     -0.026149
...
foo  std    0.966450
     min   -0.919854
     25%    0.290213
     50%    0.362949
     75%    1.247642
     max    1.548106
Length: 16, dtype: float64

The dimension of the returned result can also change:

In [104]: grouped = df.groupby('A')['C']

In [105]: def f(group):
   .....:     return DataFrame({'original' : group,
   .....:                       'demeaned' : group - group.mean()})
   .....: 

In [106]: grouped.apply(f)
Out[106]: 
   demeaned  original
0 -1.425665 -0.919854
1 -0.190202 -0.042379
2  0.741831  1.247642
3 -0.157743 -0.009920
4 -0.215598  0.290213
5  0.347944  0.495767
6 -0.142862  0.362949
7  1.042295  1.548106

apply on a Series can operate on a returned value from the applied function, that is itself a series, and possibly upcast the result to a DataFrame

In [107]: def f(x):
   .....:   return Series([ x, x**2 ], index = ['x', 'x^s'])
   .....: 

In [108]: s
Out[108]: 
first  second  third
bar    doo     one      1.346061
               two      1.511763
baz    bee     one      1.627081
               two     -0.990582
foo    bop     one     -0.441652
               two      1.211526
qux    bop     one      0.268520
               two      0.024580
dtype: float64

In [109]: s.apply(f)
Out[109]: 
                           x       x^s
first second third                    
bar   doo    one    1.346061  1.811881
             two    1.511763  2.285426
baz   bee    one    1.627081  2.647393
             two   -0.990582  0.981252
foo   bop    one   -0.441652  0.195057
             two    1.211526  1.467795
qux   bop    one    0.268520  0.072103
             two    0.024580  0.000604

Note

apply can act as a reducer, transformer, or filter function, depending on exactly what is passed to apply. So depending on the path taken, and exactly what you are grouping. Thus the grouped columns(s) may be included in the output as well as set the indices.

Warning

In the current implementation apply calls func twice on the first group to decide whether it can take a fast or slow code path. This can lead to unexpected behavior if func has side-effects, as they will take effect twice for the first group.

In [110]: d = DataFrame({"a":["x", "y"], "b":[1,2]})

In [111]: def identity(df):
   .....:     print df
   .....:     return df
   .....: 

In [112]: d.groupby("a").apply(identity)
   a  b
0  x  1
   a  b
0  x  1
   a  b
1  y  2
Out[112]: 
   a  b
0  x  1
1  y  2

Other useful features

Automatic exclusion of “nuisance” columns

Again consider the example DataFrame we’ve been looking at:

In [113]: df
Out[113]: 
     A      B         C         D
0  foo    one -0.919854 -1.131345
1  bar    one -0.042379 -0.089329
2  foo    two  1.247642  0.337863
3  bar  three -0.009920 -0.945867
4  foo    two  0.290213 -0.932132
5  bar    two  0.495767  1.956030
6  foo    one  0.362949  0.017587
7  foo  three  1.548106 -0.016692

Supposed we wished to compute the standard deviation grouped by the A column. There is a slight problem, namely that we don’t care about the data in column B. We refer to this as a “nuisance” column. If the passed aggregation function can’t be applied to some columns, the troublesome columns will be (silently) dropped. Thus, this does not pose any problems:

In [114]: df.groupby('A').std()
Out[114]: 
            C         D
A                      
bar  0.301765  1.490982
foo  0.966450  0.645875

NA group handling

If there are any NaN values in the grouping key, these will be automatically excluded. So there will never be an “NA group”. This was not the case in older versions of pandas, but users were generally discarding the NA group anyway (and supporting it was an implementation headache).

Grouping with ordered factors

Categorical variables represented as instance of pandas’s Categorical class can be used as group keys. If so, the order of the levels will be preserved:

In [115]: data = Series(np.random.randn(100))

In [116]: factor = qcut(data, [0, .25, .5, .75, 1.])

In [117]: data.groupby(factor).mean()
Out[117]: 
[-2.617, -0.684]    -1.331461
(-0.684, -0.0232]   -0.272816
(-0.0232, 0.541]     0.263607
(0.541, 2.369]       1.166038
dtype: float64

Grouping with a Grouper specification

Your may need to specify a bit more data to properly group. You can use the pd.Grouper to provide this local control.

In [118]: import datetime as DT

In [119]: df = DataFrame({
   .....:        'Branch' : 'A A A A A A A B'.split(),
   .....:        'Buyer': 'Carl Mark Carl Carl Joe Joe Joe Carl'.split(),
   .....:        'Quantity': [1,3,5,1,8,1,9,3],
   .....:        'Date' : [
   .....:              DT.datetime(2013,1,1,13,0),
   .....:              DT.datetime(2013,1,1,13,5),
   .....:              DT.datetime(2013,10,1,20,0),
   .....:              DT.datetime(2013,10,2,10,0),
   .....:              DT.datetime(2013,10,1,20,0),
   .....:              DT.datetime(2013,10,2,10,0),
   .....:              DT.datetime(2013,12,2,12,0),
   .....:              DT.datetime(2013,12,2,14,0),
   .....:              ]})
   .....: 

In [120]: df
Out[120]: 
  Branch Buyer                Date  Quantity
0      A  Carl 2013-01-01 13:00:00         1
1      A  Mark 2013-01-01 13:05:00         3
2      A  Carl 2013-10-01 20:00:00         5
3      A  Carl 2013-10-02 10:00:00         1
4      A   Joe 2013-10-01 20:00:00         8
5      A   Joe 2013-10-02 10:00:00         1
6      A   Joe 2013-12-02 12:00:00         9
7      B  Carl 2013-12-02 14:00:00         3

Groupby a specific column with the desired frequency. This is like resampling.

In [121]: df.groupby([pd.Grouper(freq='1M',key='Date'),'Buyer']).sum()
Out[121]: 
                  Quantity
Date       Buyer          
2013-01-31 Carl          1
           Mark          3
2013-10-31 Carl          6
           Joe           9
2013-12-31 Carl          3
           Joe           9

You have an ambiguous specification in that you have a named index and a column that could be potential groupers.

In [122]: df = df.set_index('Date')

In [123]: df['Date'] = df.index + pd.offsets.MonthEnd(2)

In [124]: df.groupby([pd.Grouper(freq='6M',key='Date'),'Buyer']).sum()
Out[124]: 
                  Quantity
Date       Buyer          
2013-02-28 Carl          1
           Mark          3
2014-02-28 Carl          9
           Joe          18

In [125]: df.groupby([pd.Grouper(freq='6M',level='Date'),'Buyer']).sum()
Out[125]: 
                  Quantity
Date       Buyer          
2013-01-31 Carl          1
           Mark          3
2014-01-31 Carl          9
           Joe          18

Taking the first rows of each group

Just like for a DataFrame or Series you can call head and tail on a groupby:

In [126]: df = DataFrame([[1, 2], [1, 4], [5, 6]], columns=['A', 'B'])

In [127]: df
Out[127]: 
   A  B
0  1  2
1  1  4
2  5  6

In [128]: g = df.groupby('A')

In [129]: g.head(1)
Out[129]: 
   A  B
0  1  2
2  5  6

In [130]: g.tail(1)
Out[130]: 
   A  B
1  1  4
2  5  6

This shows the first or last n rows from each group.

Warning

Before 0.14.0 this was implemented with a fall-through apply, so the result would incorrectly respect the as_index flag:

>>> g.head(1):  # was equivalent to g.apply(lambda x: x.head(1))
      A  B
 A
 1 0  1  2
 5 2  5  6

Taking the nth row of each group

To select from a DataFrame or Series the nth item, use the nth method. This is a reduction method, and will return a single row (or no row) per group:

In [131]: df = DataFrame([[1, np.nan], [1, 4], [5, 6]], columns=['A', 'B'])

In [132]: g = df.groupby('A')

In [133]: g.nth(0)
Out[133]: 
    B
A    
1 NaN
5   6

In [134]: g.nth(-1)
Out[134]: 
   B
A   
1  4
5  6

In [135]: g.nth(1)
Out[135]: 
   B
A   
1  4

If you want to select the nth not-null method, use the dropna kwarg. For a DataFrame this should be either 'any' or 'all' just like you would pass to dropna, for a Series this just needs to be truthy.

# nth(0) is the same as g.first()
In [136]: g.nth(0, dropna='any')
Out[136]: 
   B
A   
1  4
5  6

In [137]: g.first()
Out[137]: 
   B
A   
1  4
5  6

# nth(-1) is the same as g.last()
In [138]: g.nth(-1, dropna='any')  # NaNs denote group exhausted when using dropna
Out[138]: 
   B
A   
1  4
5  6

In [139]: g.last()
Out[139]: 
   B
A   
1  4
5  6

In [140]: g.B.nth(0, dropna=True)
Out[140]: 
A
1    4
5    6
Name: B, dtype: float64

As with other methods, passing as_index=False, will achieve a filtration, which returns the grouped row.

In [141]: df = DataFrame([[1, np.nan], [1, 4], [5, 6]], columns=['A', 'B'])

In [142]: g = df.groupby('A',as_index=False)

In [143]: g.nth(0)
Out[143]: 
   A   B
0  1 NaN
2  5   6

In [144]: g.nth(-1)
Out[144]: 
   A  B
1  1  4
2  5  6

Enumerate group items

New in version 0.13.0.

To see the order in which each row appears within its group, use the cumcount method:

In [145]: df = pd.DataFrame(list('aaabba'), columns=['A'])

In [146]: df
Out[146]: 
   A
0  a
1  a
2  a
3  b
4  b
5  a

In [147]: df.groupby('A').cumcount()
Out[147]: 
0    0
1    1
2    2
3    0
4    1
5    3
dtype: int64

In [148]: df.groupby('A').cumcount(ascending=False)  # kwarg only
Out[148]: 
0    3
1    2
2    1
3    1
4    0
5    0
dtype: int64

Plotting

Groupby also works with some plotting methods. For example, suppose we suspect that some features in a DataFrame my differ by group, in this case, the values in column 1 where the group is “B” are 3 higher on average.

In [149]: np.random.seed(1234)

In [150]: df = DataFrame(np.random.randn(50, 2))

In [151]: df['g'] = np.random.choice(['A', 'B'], size=50)

In [152]: df.loc[df['g'] == 'B', 1] += 3

We can easily visualize this with a boxplot:

In [153]: df.groupby('g').boxplot()
Out[153]: OrderedDict([('A', {'medians': [<matplotlib.lines.Line2D object at 0xa2ee5d6c>, <matplotlib.lines.Line2D object at 0xa3624a2c>], 'fliers': [<matplotlib.lines.Line2D object at 0xa2ee544c>, <matplotlib.lines.Line2D object at 0xa2ee50cc>, <matplotlib.lines.Line2D object at 0xa3624acc>, <matplotlib.lines.Line2D object at 0xa3a21c2c>], 'whiskers': [<matplotlib.lines.Line2D object at 0xa352ae2c>, <matplotlib.lines.Line2D object at 0xa352a50c>, <matplotlib.lines.Line2D object at 0xa2ec73ac>, <matplotlib.lines.Line2D object at 0xa2ec8fac>], 'boxes': [<matplotlib.lines.Line2D object at 0xa2eadf2c>, <matplotlib.lines.Line2D object at 0xa2d496ec>], 'caps': [<matplotlib.lines.Line2D object at 0xa2eade8c>, <matplotlib.lines.Line2D object at 0xa2ead86c>, <matplotlib.lines.Line2D object at 0xa3687f6c>, <matplotlib.lines.Line2D object at 0xadb8a94c>]}), ('B', {'medians': [<matplotlib.lines.Line2D object at 0xa2ebd04c>, <matplotlib.lines.Line2D object at 0xa317d5cc>], 'fliers': [<matplotlib.lines.Line2D object at 0xa35b2c2c>, <matplotlib.lines.Line2D object at 0xa2d09e4c>, <matplotlib.lines.Line2D object at 0xa2d0970c>, <matplotlib.lines.Line2D object at 0xa299604c>], 'whiskers': [<matplotlib.lines.Line2D object at 0xa2d2396c>, <matplotlib.lines.Line2D object at 0xa296e58c>, <matplotlib.lines.Line2D object at 0xa2eb12ac>, <matplotlib.lines.Line2D object at 0xa2eb222c>], 'boxes': [<matplotlib.lines.Line2D object at 0xa2ebd36c>, <matplotlib.lines.Line2D object at 0xa297bb4c>], 'caps': [<matplotlib.lines.Line2D object at 0xa2c1392c>, <matplotlib.lines.Line2D object at 0xa2c1344c>, <matplotlib.lines.Line2D object at 0xa297378c>, <matplotlib.lines.Line2D object at 0xa2975b2c>]})])
_images/groupby_boxplot.png

The result of calling boxplot is a dictionary whose keys are the values of our grouping column g (“A” and “B”). The values of the resulting dictionary can be controlled by the return_type keyword of boxplot. See the visualization documentation for more.

Warning

For historical reasons, df.groupby("g").boxplot() is not equivalent to df.boxplot(by="g"). See here for an explanation.

Examples

Regrouping by factor

Regroup columns of a DataFrame according to their sum, and sum the aggregated ones.

In [154]: df = pd.DataFrame({'a':[1,0,0], 'b':[0,1,0], 'c':[1,0,0], 'd':[2,3,4]})

In [155]: df
Out[155]: 
   a  b  c  d
0  1  0  1  2
1  0  1  0  3
2  0  0  0  4

In [156]: df.groupby(df.sum(), axis=1).sum()
Out[156]: 
   1  9
0  2  2
1  1  3
2  0  4

Returning a Series to propogate names

Group DataFrame columns, compute a set of metrics and return a named Series. The Series name is used as the name for the column index. This is especially useful in conjunction with reshaping operations such as stacking in which the column index name will be used as the name of the inserted column:

In [157]: df = pd.DataFrame({
   .....:      'a':  [0, 0, 0, 0, 1, 1, 1, 1, 2, 2, 2, 2],
   .....:      'b':  [0, 0, 1, 1, 0, 0, 1, 1, 0, 0, 1, 1],
   .....:      'c':  [1, 0, 1, 0, 1, 0, 1, 0, 1, 0, 1, 0],
   .....:      'd':  [0, 0, 0, 1, 0, 0, 0, 1, 0, 0, 0, 1],
   .....:      })
   .....: 

In [158]: def compute_metrics(x):
   .....:     result = {'b_sum': x['b'].sum(), 'c_mean': x['c'].mean()}
   .....:     return pd.Series(result, name='metrics')
   .....: 

In [159]: result = df.groupby('a').apply(compute_metrics)

In [160]: result
Out[160]: 
metrics  b_sum  c_mean
a                     
0            2     0.5
1            2     0.5
2            2     0.5

In [161]: result.stack()
Out[161]: 
a  metrics
0  b_sum      2.0
   c_mean     0.5
1  b_sum      2.0
   c_mean     0.5
2  b_sum      2.0
   c_mean     0.5
dtype: float64