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.
Out of these, the split step is the most straightforward. In fact, in many situations we may wish to split the data set into groups and do something with those groups. In the apply step, we might wish to do one of the following:
Aggregation: compute a summary statistic (or statistics) for each group. Some examples:
Compute group sums or means.
Compute group sizes / counts.
Transformation: perform some group-specific computations and return a like-indexed object. Some examples:
Standardize data (zscore) within a group.
Filling NAs within groups with a value derived from each group.
Filtration: discard some groups, according to a group-wise computation that evaluates to True or False. Some examples:
Discard data that belong to groups with only a few members.
Filter out data based on the group sum or mean.
Many of these operations are defined on GroupBy objects. These operations are similar to those of the aggregating API, window API, and resample API.
It is possible that a given operation does not fall into one of these categories or
is some combination of them. In such a case, it may be possible to compute the
operation using GroupBy’s apply method. This method will examine the results of the
apply step and try to sensibly combine them into a single result if it doesn’t fit into either
of the above three categories.
Note
An operation that is split into multiple steps using built-in GroupBy operations
will be more efficient than using the apply method with a user-defined Python
function.
Since the set of object instance methods on pandas data structures is 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#
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 may do the following:
In [1]: speeds = pd.DataFrame(
...: [
...: ("bird", "Falconiformes", 389.0),
...: ("bird", "Psittaciformes", 24.0),
...: ("mammal", "Carnivora", 80.2),
...: ("mammal", "Primates", np.nan),
...: ("mammal", "Carnivora", 58),
...: ],
...: index=["falcon", "parrot", "lion", "monkey", "leopard"],
...: columns=("class", "order", "max_speed"),
...: )
...:
In [2]: speeds
Out[2]:
class order max_speed
falcon bird Falconiformes 389.0
parrot bird Psittaciformes 24.0
lion mammal Carnivora 80.2
monkey mammal Primates NaN
leopard mammal Carnivora 58.0
In [3]: grouped = speeds.groupby("class")
In [4]: grouped = speeds.groupby(["class", "order"])
The mapping can be specified many different ways:
A Python function, to be called on each of the index labels.
A list or NumPy array of the same length as the index.
A dict or
Series, providing alabel -> group namemapping.For
DataFrameobjects, a string indicating either a column name or an index level name to be used to group.A list of any of the above things.
Collectively we refer to the grouping objects as the keys. For example,
consider the following DataFrame:
Note
A string passed to groupby may refer to either a column or an index level.
If a string matches both a column name and an index level name, a
ValueError will be raised.
In [5]: df = pd.DataFrame(
...: {
...: "A": ["foo", "bar", "foo", "bar", "foo", "bar", "foo", "foo"],
...: "B": ["one", "one", "two", "three", "two", "two", "one", "three"],
...: "C": np.random.randn(8),
...: "D": np.random.randn(8),
...: }
...: )
...:
In [6]: df
Out[6]:
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
On a DataFrame, we obtain a GroupBy object by calling groupby().
This method returns a pandas.api.typing.DataFrameGroupBy instance.
We could naturally group by either the A or B columns, or both:
In [7]: grouped = df.groupby("A")
In [8]: grouped = df.groupby(["A", "B"])
Note
df.groupby('A') is just syntactic sugar for df.groupby(df['A']).
If we also have a MultiIndex on columns A and B, we can group by all
the columns except the one we specify:
In [9]: df2 = df.set_index(["A", "B"])
In [10]: grouped = df2.groupby(level=df2.index.names.difference(["B"]))
In [11]: grouped.sum()
Out[11]:
C D
A
bar -1.591710 -1.739537
foo -0.752861 -1.402938
The above GroupBy will split the DataFrame on its index (rows). To split by columns, first do a transpose:
In [12]: def get_letter_type(letter):
....: if letter.lower() in 'aeiou':
....: return 'vowel'
....: else:
....: return 'consonant'
....:
In [13]: grouped = df.T.groupby(get_letter_type)
pandas Index objects support 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 [14]: lst = [1, 2, 3, 1, 2, 3]
In [15]: s = pd.Series([1, 2, 3, 10, 20, 30], lst)
In [16]: grouped = s.groupby(level=0)
In [17]: grouped.first()
Out[17]:
1 1
2 2
3 3
dtype: int64
In [18]: grouped.last()
Out[18]:
1 10
2 20
3 30
dtype: int64
In [19]: grouped.sum()
Out[19]:
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 it can’t be guaranteed to be the most efficient implementation). You can get quite creative with the label mapping functions.
GroupBy sorting#
By default the group keys are sorted during the groupby operation. You may however pass sort=False for potential speedups. With sort=False the order among group-keys follows the order of appearance of the keys in the original dataframe:
In [20]: df2 = pd.DataFrame({"X": ["B", "B", "A", "A"], "Y": [1, 2, 3, 4]})
In [21]: df2.groupby(["X"]).sum()
Out[21]:
Y
X
A 7
B 3
In [22]: df2.groupby(["X"], sort=False).sum()
Out[22]:
Y
X
B 3
A 7
Note that groupby will preserve the order in which observations are sorted within each group.
For example, the groups created by groupby() below are in the order they appeared in the original DataFrame:
In [23]: df3 = pd.DataFrame({"X": ["A", "B", "A", "B"], "Y": [1, 4, 3, 2]})
In [24]: df3.groupby(["X"]).get_group("A")
Out[24]:
X Y
0 A 1
2 A 3
In [25]: df3.groupby(["X"]).get_group("B")
Out[25]:
X Y
1 B 4
3 B 2
GroupBy dropna#
By default NA values are excluded from group keys during the groupby operation. However,
in case you want to include NA values in group keys, you could pass dropna=False to achieve it.
In [26]: df_list = [[1, 2, 3], [1, None, 4], [2, 1, 3], [1, 2, 2]]
In [27]: df_dropna = pd.DataFrame(df_list, columns=["a", "b", "c"])
In [28]: df_dropna
Out[28]:
a b c
0 1 2.0 3
1 1 NaN 4
2 2 1.0 3
3 1 2.0 2
# Default ``dropna`` is set to True, which will exclude NaNs in keys
In [29]: df_dropna.groupby(by=["b"], dropna=True).sum()
Out[29]:
a c
b
1.0 2 3
2.0 2 5
# In order to allow NaN in keys, set ``dropna`` to False
In [30]: df_dropna.groupby(by=["b"], dropna=False).sum()
Out[30]:
a c
b
1.0 2 3
2.0 2 5
NaN 1 4
The default setting of dropna argument is True which means NA are not included in group keys.
GroupBy object attributes#
The groups attribute is a dictionary whose keys are the computed unique groups
and corresponding values are the axis labels belonging to each group. In the
above example we have:
In [31]: df.groupby("A").groups
Out[31]: {'bar': [1, 3, 5], 'foo': [0, 2, 4, 6, 7]}
In [32]: df.T.groupby(get_letter_type).groups
Out[32]: {'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 [33]: grouped = df.groupby(["A", "B"])
In [34]: grouped.groups
Out[34]: {('bar', 'one'): [1], ('bar', 'three'): [3], ('bar', 'two'): [5], ('foo', 'one'): [0, 6], ('foo', 'three'): [7], ('foo', 'two'): [2, 4]}
In [35]: len(grouped)
Out[35]: 6
GroupBy will tab complete column names (and other attributes):
In [36]: n = 10
In [37]: weight = np.random.normal(166, 20, size=n)
In [38]: height = np.random.normal(60, 10, size=n)
In [39]: time = pd.date_range("1/1/2000", periods=n)
In [40]: gender = np.random.choice(["male", "female"], size=n)
In [41]: df = pd.DataFrame(
....: {"height": height, "weight": weight, "gender": gender}, index=time
....: )
....:
In [42]: df
Out[42]:
height weight gender
2000-01-01 42.849980 157.500553 male
2000-01-02 49.607315 177.340407 male
2000-01-03 56.293531 171.524640 male
2000-01-04 48.421077 144.251986 female
2000-01-05 46.556882 152.526206 male
2000-01-06 68.448851 168.272968 female
2000-01-07 70.757698 136.431469 male
2000-01-08 58.909500 176.499753 female
2000-01-09 76.435631 174.094104 female
2000-01-10 45.306120 177.540920 male
In [43]: gb = df.groupby("gender")
In [44]: gb.<TAB> # noqa: E225, E999
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.
Let’s create a Series with a two-level MultiIndex.
In [45]: arrays = [
....: ["bar", "bar", "baz", "baz", "foo", "foo", "qux", "qux"],
....: ["one", "two", "one", "two", "one", "two", "one", "two"],
....: ]
....:
In [46]: index = pd.MultiIndex.from_arrays(arrays, names=["first", "second"])
In [47]: s = pd.Series(np.random.randn(8), index=index)
In [48]: s
Out[48]:
first second
bar one -0.919854
two -0.042379
baz one 1.247642
two -0.009920
foo one 0.290213
two 0.495767
qux one 0.362949
two 1.548106
dtype: float64
We can then group by one of the levels in s.
In [49]: grouped = s.groupby(level=0)
In [50]: grouped.sum()
Out[50]:
first
bar -0.962232
baz 1.237723
foo 0.785980
qux 1.911055
dtype: float64
If the MultiIndex has names specified, these can be passed instead of the level number:
In [51]: s.groupby(level="second").sum()
Out[51]:
second
one 0.980950
two 1.991575
dtype: float64
Grouping with multiple levels is supported.
In [52]: arrays = [
....: ["bar", "bar", "baz", "baz", "foo", "foo", "qux", "qux"],
....: ["doo", "doo", "bee", "bee", "bop", "bop", "bop", "bop"],
....: ["one", "two", "one", "two", "one", "two", "one", "two"],
....: ]
....:
In [53]: index = pd.MultiIndex.from_arrays(arrays, names=["first", "second", "third"])
In [54]: s = pd.Series(np.random.randn(8), index=index)
In [55]: s
Out[55]:
first second third
bar doo one -1.131345
two -0.089329
baz bee one 0.337863
two -0.945867
foo bop one -0.932132
two 1.956030
qux bop one 0.017587
two -0.016692
dtype: float64
In [56]: s.groupby(level=["first", "second"]).sum()
Out[56]:
first second
bar doo -1.220674
baz bee -0.608004
foo bop 1.023898
qux bop 0.000895
dtype: float64
Index level names may be supplied as keys.
In [57]: s.groupby(["first", "second"]).sum()
Out[57]:
first second
bar doo -1.220674
baz bee -0.608004
foo bop 1.023898
qux bop 0.000895
dtype: float64
More on the sum function and aggregation later.
Grouping DataFrame with Index levels and columns#
A DataFrame may be grouped by a combination of columns and index levels. You
can specify both column and index names, or use a Grouper.
Let’s first create a DataFrame with a MultiIndex:
In [58]: arrays = [
....: ["bar", "bar", "baz", "baz", "foo", "foo", "qux", "qux"],
....: ["one", "two", "one", "two", "one", "two", "one", "two"],
....: ]
....:
In [59]: index = pd.MultiIndex.from_arrays(arrays, names=["first", "second"])
In [60]: df = pd.DataFrame({"A": [1, 1, 1, 1, 2, 2, 3, 3], "B": np.arange(8)}, index=index)
In [61]: df
Out[61]:
A B
first second
bar one 1 0
two 1 1
baz one 1 2
two 1 3
foo one 2 4
two 2 5
qux one 3 6
two 3 7
Then we group df by the second index level and the A column.
In [62]: df.groupby([pd.Grouper(level=1), "A"]).sum()
Out[62]:
B
second A
one 1 2
2 4
3 6
two 1 4
2 5
3 7
Index levels may also be specified by name.
In [63]: df.groupby([pd.Grouper(level="second"), "A"]).sum()
Out[63]:
B
second A
one 1 2
2 4
3 6
two 1 4
2 5
3 7
Index level names may be specified as keys directly to groupby.
In [64]: df.groupby(["second", "A"]).sum()
Out[64]:
B
second A
one 1 2
2 4
3 6
two 1 4
2 5
3 7
DataFrame column selection in GroupBy#
Once you have created the GroupBy object from a DataFrame, you might want to do
something different for each of the columns. Thus, by using [] on the GroupBy
object in a similar way as the one used to get a column from a DataFrame, you can do:
In [65]: df = pd.DataFrame(
....: {
....: "A": ["foo", "bar", "foo", "bar", "foo", "bar", "foo", "foo"],
....: "B": ["one", "one", "two", "three", "two", "two", "one", "three"],
....: "C": np.random.randn(8),
....: "D": np.random.randn(8),
....: }
....: )
....:
In [66]: df
Out[66]:
A B C D
0 foo one -0.575247 1.346061
1 bar one 0.254161 1.511763
2 foo two -1.143704 1.627081
3 bar three 0.215897 -0.990582
4 foo two 1.193555 -0.441652
5 bar two -0.077118 1.211526
6 foo one -0.408530 0.268520
7 foo three -0.862495 0.024580
In [67]: grouped = df.groupby(["A"])
In [68]: grouped_C = grouped["C"]
In [69]: grouped_D = grouped["D"]
This is mainly syntactic sugar for the alternative, which is much more verbose:
In [70]: df["C"].groupby(df["A"])
Out[70]: <pandas.core.groupby.generic.SeriesGroupBy object at 0x7fd7dc57c7f0>
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 [71]: grouped = df.groupby('A')
In [72]: for name, group in grouped:
....: print(name)
....: print(group)
....:
bar
A B C D
1 bar one 0.254161 1.511763
3 bar three 0.215897 -0.990582
5 bar two -0.077118 1.211526
foo
A B C D
0 foo one -0.575247 1.346061
2 foo two -1.143704 1.627081
4 foo two 1.193555 -0.441652
6 foo one -0.408530 0.268520
7 foo three -0.862495 0.024580
In the case of grouping by multiple keys, the group name will be a tuple:
In [73]: for name, group in df.groupby(['A', 'B']):
....: print(name)
....: print(group)
....:
('bar', 'one')
A B C D
1 bar one 0.254161 1.511763
('bar', 'three')
A B C D
3 bar three 0.215897 -0.990582
('bar', 'two')
A B C D
5 bar two -0.077118 1.211526
('foo', 'one')
A B C D
0 foo one -0.575247 1.346061
6 foo one -0.408530 0.268520
('foo', 'three')
A B C D
7 foo three -0.862495 0.02458
('foo', 'two')
A B C D
2 foo two -1.143704 1.627081
4 foo two 1.193555 -0.441652
Selecting a group#
A single group can be selected using
get_group():
In [74]: grouped.get_group("bar")
Out[74]:
A B C D
1 bar one 0.254161 1.511763
3 bar three 0.215897 -0.990582
5 bar two -0.077118 1.211526
Or for an object grouped on multiple columns:
In [75]: df.groupby(["A", "B"]).get_group(("bar", "one"))
Out[75]:
A B C D
1 bar one 0.254161 1.511763
Aggregation#
An aggregation is a GroupBy operation that reduces the dimension of the grouping object. The result of an aggregation is, or at least is treated as, a scalar value for each column in a group. For example, producing the sum of each column in a group of values.
In [76]: animals = 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 [77]: animals
Out[77]:
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 [78]: animals.groupby("kind").sum()
Out[78]:
height weight
kind
cat 18.6 17.8
dog 40.0 205.5
In the result, the keys of the groups appear in the index by default. They can be
instead included in the columns by passing as_index=False.
In [79]: animals.groupby("kind", as_index=False).sum()
Out[79]:
kind height weight
0 cat 18.6 17.8
1 dog 40.0 205.5
Built-in aggregation methods#
Many common aggregations are built-in to GroupBy objects as methods. Of the methods
listed below, those with a * do not have a Cython-optimized implementation.
Method |
Description |
|---|---|
Compute whether any of the values in the groups are truthy |
|
Compute whether all of the values in the groups are truthy |
|
Compute the number of non-NA values in the groups |
|
|
Compute the covariance of the groups |
Compute the first occurring value in each group |
|
|
Compute the index of the maximum value in each group |
|
Compute the index of the minimum value in each group |
Compute the last occurring value in each group |
|
Compute the maximum value in each group |
|
Compute the mean of each group |
|
Compute the median of each group |
|
Compute the minimum value in each group |
|
Compute the number of unique values in each group |
|
Compute the product of the values in each group |
|
Compute a given quantile of the values in each group |
|
Compute the standard error of the mean of the values in each group |
|
Compute the number of values in each group |
|
|
Compute the skew of the values in each group |
Compute the standard deviation of the values in each group |
|
Compute the sum of the values in each group |
|
Compute the variance of the values in each group |
Some examples:
In [80]: df.groupby("A")[["C", "D"]].max()
Out[80]:
C D
A
bar 0.254161 1.511763
foo 1.193555 1.627081
In [81]: df.groupby(["A", "B"]).mean()
Out[81]:
C D
A B
bar one 0.254161 1.511763
three 0.215897 -0.990582
two -0.077118 1.211526
foo one -0.491888 0.807291
three -0.862495 0.024580
two 0.024925 0.592714
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 [82]: grouped = df.groupby(["A", "B"])
In [83]: grouped.size()
Out[83]:
A B
bar one 1
three 1
two 1
foo one 2
three 1
two 2
dtype: int64
While the describe() method is not itself a reducer, it
can be used to conveniently produce a collection of summary statistics about each of
the groups.
In [84]: grouped.describe()
Out[84]:
C ... D
count mean std ... 50% 75% max
A B ...
bar one 1.0 0.254161 NaN ... 1.511763 1.511763 1.511763
three 1.0 0.215897 NaN ... -0.990582 -0.990582 -0.990582
two 1.0 -0.077118 NaN ... 1.211526 1.211526 1.211526
foo one 2.0 -0.491888 0.117887 ... 0.807291 1.076676 1.346061
three 1.0 -0.862495 NaN ... 0.024580 0.024580 0.024580
two 2.0 0.024925 1.652692 ... 0.592714 1.109898 1.627081
[6 rows x 16 columns]
Another aggregation example is to compute the number of unique values of each group.
This is similar to the value_counts function, except that it only counts the
number of unique values.
In [85]: ll = [['foo', 1], ['foo', 2], ['foo', 2], ['bar', 1], ['bar', 1]]
In [86]: df4 = pd.DataFrame(ll, columns=["A", "B"])
In [87]: df4
Out[87]:
A B
0 foo 1
1 foo 2
2 foo 2
3 bar 1
4 bar 1
In [88]: df4.groupby("A")["B"].nunique()
Out[88]:
A
bar 1
foo 2
Name: B, dtype: int64
Note
Aggregation functions will not return the groups that you are aggregating over
as 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 indices or columns.
The aggregate() method#
Note
The aggregate() method can accept many different types of
inputs. This section details using string aliases for various GroupBy methods; other
inputs are detailed in the sections below.
Any reduction method that pandas implements can be passed as a string to
aggregate(). Users are encouraged to use the shorthand,
agg. It will operate as if the corresponding method was called.
In [89]: grouped = df.groupby("A")
In [90]: grouped[["C", "D"]].aggregate("sum")
Out[90]:
C D
A
bar 0.392940 1.732707
foo -1.796421 2.824590
In [91]: grouped = df.groupby(["A", "B"])
In [92]: grouped.agg("sum")
Out[92]:
C D
A B
bar one 0.254161 1.511763
three 0.215897 -0.990582
two -0.077118 1.211526
foo one -0.983776 1.614581
three -0.862495 0.024580
two 0.049851 1.185429
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. As mentioned above, this can be
changed by using the as_index option:
In [93]: grouped = df.groupby(["A", "B"], as_index=False)
In [94]: grouped.agg("sum")
Out[94]:
A B C D
0 bar one 0.254161 1.511763
1 bar three 0.215897 -0.990582
2 bar two -0.077118 1.211526
3 foo one -0.983776 1.614581
4 foo three -0.862495 0.024580
5 foo two 0.049851 1.185429
In [95]: df.groupby("A", as_index=False)[["C", "D"]].agg("sum")
Out[95]:
A C D
0 bar 0.392940 1.732707
1 foo -1.796421 2.824590
Note that you could use the DataFrame.reset_index() DataFrame function to achieve
the same result as the column names are stored in the resulting MultiIndex, although
this will make an extra copy.
In [96]: df.groupby(["A", "B"]).agg("sum").reset_index()
Out[96]:
A B C D
0 bar one 0.254161 1.511763
1 bar three 0.215897 -0.990582
2 bar two -0.077118 1.211526
3 foo one -0.983776 1.614581
4 foo three -0.862495 0.024580
5 foo two 0.049851 1.185429
Aggregation with User-Defined Functions#
Users can also provide their own User-Defined Functions (UDFs) for custom aggregations.
Warning
When aggregating with a UDF, the UDF should not mutate the
provided Series. See Mutating with User Defined Function (UDF) methods for more information.
Note
Aggregating with a UDF is often less performant than using the pandas built-in methods on GroupBy. Consider breaking up a complex operation into a chain of operations that utilize the built-in methods.
In [97]: animals
Out[97]:
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 [98]: animals.groupby("kind")[["height"]].agg(lambda x: set(x))
Out[98]:
height
kind
cat {9.1, 9.5}
dog {34.0, 6.0}
The resulting dtype will reflect that of the aggregating function. If the results from different groups have
different dtypes, then a common dtype will be determined in the same way as DataFrame construction.
In [99]: animals.groupby("kind")[["height"]].agg(lambda x: x.astype(int).sum())
Out[99]:
height
kind
cat 18
dog 40
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 [100]: grouped = df.groupby("A")
In [101]: grouped["C"].agg(["sum", "mean", "std"])
Out[101]:
sum mean std
A
bar 0.392940 0.130980 0.181231
foo -1.796421 -0.359284 0.912265
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 [102]: grouped[["C", "D"]].agg(["sum", "mean", "std"])
Out[102]:
C D
sum mean std sum mean std
A
bar 0.392940 0.130980 0.181231 1.732707 0.577569 1.366330
foo -1.796421 -0.359284 0.912265 2.824590 0.564918 0.884785
The resulting aggregations are named after the functions themselves. If you
need to rename, then you can add in a chained operation for a Series like this:
In [103]: (
.....: grouped["C"]
.....: .agg(["sum", "mean", "std"])
.....: .rename(columns={"sum": "foo", "mean": "bar", "std": "baz"})
.....: )
.....:
Out[103]:
foo bar baz
A
bar 0.392940 0.130980 0.181231
foo -1.796421 -0.359284 0.912265
For a grouped DataFrame, you can rename in a similar manner:
In [104]: (
.....: grouped[["C", "D"]].agg(["sum", "mean", "std"]).rename(
.....: columns={"sum": "foo", "mean": "bar", "std": "baz"}
.....: )
.....: )
.....:
Out[104]:
C D
foo bar baz foo bar baz
A
bar 0.392940 0.130980 0.181231 1.732707 0.577569 1.366330
foo -1.796421 -0.359284 0.912265 2.824590 0.564918 0.884785
Note
In general, the output column names should be unique, but pandas will allow you apply to the same function (or two functions with the same name) to the same column.
In [105]: grouped["C"].agg(["sum", "sum"])
Out[105]:
sum sum
A
bar 0.392940 0.392940
foo -1.796421 -1.796421
pandas also allows you to provide multiple lambdas. In this case, pandas
will mangle the name of the (nameless) lambda functions, appending _<i>
to each subsequent lambda.
In [106]: grouped["C"].agg([lambda x: x.max() - x.min(), lambda x: x.median() - x.mean()])
Out[106]:
<lambda_0> <lambda_1>
A
bar 0.331279 0.084917
foo 2.337259 -0.215962
Named aggregation#
To support column-specific aggregation with control over the output column names, pandas
accepts the special syntax in DataFrameGroupBy.agg() and SeriesGroupBy.agg(), known as “named aggregation”, where
The keywords are the output column names
The values are tuples whose first element is the column to select and the second element is the aggregation to apply to that column. pandas provides the
NamedAggnamedtuple with the fields['column', 'aggfunc']to make it clearer what the arguments are. As usual, the aggregation can be a callable or a string alias.
In [107]: animals
Out[107]:
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 [108]: animals.groupby("kind").agg(
.....: min_height=pd.NamedAgg(column="height", aggfunc="min"),
.....: max_height=pd.NamedAgg(column="height", aggfunc="max"),
.....: average_weight=pd.NamedAgg(column="weight", aggfunc="mean"),
.....: )
.....:
Out[108]:
min_height max_height average_weight
kind
cat 9.1 9.5 8.90
dog 6.0 34.0 102.75
NamedAgg is just a namedtuple. Plain tuples are allowed as well.
In [109]: animals.groupby("kind").agg(
.....: min_height=("height", "min"),
.....: max_height=("height", "max"),
.....: average_weight=("weight", "mean"),
.....: )
.....:
Out[109]:
min_height max_height average_weight
kind
cat 9.1 9.5 8.90
dog 6.0 34.0 102.75
If the column names you want are not valid Python keywords, construct a dictionary and unpack the keyword arguments
In [110]: animals.groupby("kind").agg(
.....: **{
.....: "total weight": pd.NamedAgg(column="weight", aggfunc="sum")
.....: }
.....: )
.....:
Out[110]:
total weight
kind
cat 17.8
dog 205.5
When using named aggregation, additional keyword arguments are not passed through
to the aggregation functions; only pairs
of (column, aggfunc) should be passed as **kwargs. If your aggregation functions
require additional arguments, apply them partially with functools.partial().
Named aggregation is also valid for Series groupby aggregations. In this case there’s no column selection, so the values are just the functions.
In [111]: animals.groupby("kind").height.agg(
.....: min_height="min",
.....: max_height="max",
.....: )
.....:
Out[111]:
min_height max_height
kind
cat 9.1 9.5
dog 6.0 34.0
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 [112]: grouped.agg({"C": "sum", "D": lambda x: np.std(x, ddof=1)})
Out[112]:
C D
A
bar 0.392940 1.366330
foo -1.796421 0.884785
The function names can also be strings. In order for a string to be valid it must be implemented on GroupBy:
In [113]: grouped.agg({"C": "sum", "D": "std"})
Out[113]:
C D
A
bar 0.392940 1.366330
foo -1.796421 0.884785
Transformation#
A transformation is a GroupBy operation whose result is indexed the same
as the one being grouped. Common examples include cumsum() and
diff().
In [114]: speeds
Out[114]:
class order max_speed
falcon bird Falconiformes 389.0
parrot bird Psittaciformes 24.0
lion mammal Carnivora 80.2
monkey mammal Primates NaN
leopard mammal Carnivora 58.0
In [115]: grouped = speeds.groupby("class")["max_speed"]
In [116]: grouped.cumsum()
Out[116]:
falcon 389.0
parrot 413.0
lion 80.2
monkey NaN
leopard 138.2
Name: max_speed, dtype: float64
In [117]: grouped.diff()
Out[117]:
falcon NaN
parrot -365.0
lion NaN
monkey NaN
leopard NaN
Name: max_speed, dtype: float64
Unlike aggregations, the groupings that are used to split the original object are not included in the result.
Note
Since transformations do not include the groupings that are used to split the result,
the arguments as_index and sort in DataFrame.groupby() and
Series.groupby() have no effect.
A common use of a transformation is to add the result back into the original DataFrame.
In [118]: result = speeds.copy()
In [119]: result["cumsum"] = grouped.cumsum()
In [120]: result["diff"] = grouped.diff()
In [121]: result
Out[121]:
class order max_speed cumsum diff
falcon bird Falconiformes 389.0 389.0 NaN
parrot bird Psittaciformes 24.0 413.0 -365.0
lion mammal Carnivora 80.2 80.2 NaN
monkey mammal Primates NaN NaN NaN
leopard mammal Carnivora 58.0 138.2 NaN
Built-in transformation methods#
The following methods on GroupBy act as transformations. Of these methods, only
fillna does not have a Cython-optimized implementation.
Method |
Description |
|---|---|
Back fill NA values within each group |
|
Compute the cumulative count within each group |
|
Compute the cumulative max within each group |
|
Compute the cumulative min within each group |
|
Compute the cumulative product within each group |
|
Compute the cumulative sum within each group |
|
Compute the difference between adjacent values within each group |
|
Forward fill NA values within each group |
|
Fill NA values within each group |
|
Compute the percent change between adjacent values within each group |
|
Compute the rank of each value within each group |
|
Shift values up or down within each group |
In addition, passing any built-in aggregation method as a string to
transform() (see the next section) will broadcast the result
across the group, producing a transformed result. If the aggregation method is
Cython-optimized, this will be performant as well.
The transform() method#
Similar to the aggregation method, the
transform() method can accept string aliases to the built-in
transformation methods in the previous section. It can also accept string aliases to
the built-in aggregation methods. When an aggregation method is provided, the result
will be broadcast across the group.
In [122]: speeds
Out[122]:
class order max_speed
falcon bird Falconiformes 389.0
parrot bird Psittaciformes 24.0
lion mammal Carnivora 80.2
monkey mammal Primates NaN
leopard mammal Carnivora 58.0
In [123]: grouped = speeds.groupby("class")[["max_speed"]]
In [124]: grouped.transform("cumsum")
Out[124]:
max_speed
falcon 389.0
parrot 413.0
lion 80.2
monkey NaN
leopard 138.2
In [125]: grouped.transform("sum")
Out[125]:
max_speed
falcon 413.0
parrot 413.0
lion 138.2
monkey 138.2
leopard 138.2
In addition to string aliases, the transform() method can
also accept User-Defined Functions (UDFs). The UDF must:
Return a result that is either the same size as the group chunk or broadcastable to the size of the group chunk (e.g., a scalar,
grouped.transform(lambda x: x.iloc[-1])).Operate column-by-column on the group chunk. The transform is applied to the first group chunk using chunk.apply.
Not perform in-place operations on the group chunk. Group chunks should be treated as immutable, and changes to a group chunk may produce unexpected results. See Mutating with User Defined Function (UDF) methods for more information.
(Optionally) operates on all columns of the entire group chunk at once. If this is supported, a fast path is used starting from the second chunk.
Note
Transforming by supplying transform with a UDF is
often less performant than using the built-in methods on GroupBy.
Consider breaking up a complex operation into a chain of operations that utilize
the built-in methods.
All of the examples in this section can be made more performant by calling
built-in methods instead of using transform.
See below for examples.
Changed in version 2.0.0: When using .transform on a grouped DataFrame and the transformation function
returns a DataFrame, pandas now aligns the result’s index
with the input’s index. You can call .to_numpy() within the transformation
function to avoid alignment.
Similar to The aggregate() method, the resulting dtype will reflect that of the
transformation function. If the results from different groups have different dtypes, then
a common dtype will be determined in the same way as DataFrame construction.
Suppose we wish to standardize the data within each group:
In [126]: index = pd.date_range("10/1/1999", periods=1100)
In [127]: ts = pd.Series(np.random.normal(0.5, 2, 1100), index)
In [128]: ts = ts.rolling(window=100, min_periods=100).mean().dropna()
In [129]: ts.head()
Out[129]:
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 [130]: ts.tail()
Out[130]:
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 [131]: transformed = ts.groupby(lambda x: x.year).transform(
.....: lambda x: (x - x.mean()) / x.std()
.....: )
.....:
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 [132]: grouped = ts.groupby(lambda x: x.year)
In [133]: grouped.mean()
Out[133]:
2000 0.442441
2001 0.526246
2002 0.459365
dtype: float64
In [134]: grouped.std()
Out[134]:
2000 0.131752
2001 0.210945
2002 0.128753
dtype: float64
# Transformed Data
In [135]: grouped_trans = transformed.groupby(lambda x: x.year)
In [136]: grouped_trans.mean()
Out[136]:
2000 -4.870756e-16
2001 -1.545187e-16
2002 4.136282e-16
dtype: float64
In [137]: grouped_trans.std()
Out[137]:
2000 1.0
2001 1.0
2002 1.0
dtype: float64
We can also visually compare the original and transformed data sets.
In [138]: compare = pd.DataFrame({"Original": ts, "Transformed": transformed})
In [139]: compare.plot()
Out[139]: <Axes: >
Transformation functions that have lower dimension outputs are broadcast to match the shape of the input array.
In [140]: ts.groupby(lambda x: x.year).transform(lambda x: x.max() - x.min())
Out[140]:
2000-01-08 0.623893
2000-01-09 0.623893
2000-01-10 0.623893
2000-01-11 0.623893
2000-01-12 0.623893
...
2002-09-30 0.558275
2002-10-01 0.558275
2002-10-02 0.558275
2002-10-03 0.558275
2002-10-04 0.558275
Freq: D, Length: 1001, dtype: float64
Another common data transform is to replace missing data with the group mean.
In [141]: cols = ["A", "B", "C"]
In [142]: values = np.random.randn(1000, 3)
In [143]: values[np.random.randint(0, 1000, 100), 0] = np.nan
In [144]: values[np.random.randint(0, 1000, 50), 1] = np.nan
In [145]: values[np.random.randint(0, 1000, 200), 2] = np.nan
In [146]: data_df = pd.DataFrame(values, columns=cols)
In [147]: data_df
Out[147]:
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
.. ... ... ...
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 [148]: countries = np.array(["US", "UK", "GR", "JP"])
In [149]: key = countries[np.random.randint(0, 4, 1000)]
In [150]: grouped = data_df.groupby(key)
# Non-NA count in each group
In [151]: grouped.count()
Out[151]:
A B C
GR 209 217 189
JP 240 255 217
UK 216 231 193
US 239 250 217
In [152]: transformed = grouped.transform(lambda x: x.fillna(x.mean()))
We can verify that the group means have not changed in the transformed data, and that the transformed data contains no NAs.
In [153]: grouped_trans = transformed.groupby(key)
In [154]: grouped.mean() # original group means
Out[154]:
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 [155]: grouped_trans.mean() # transformation did not change group means
Out[155]:
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 [156]: grouped.count() # original has some missing data points
Out[156]:
A B C
GR 209 217 189
JP 240 255 217
UK 216 231 193
US 239 250 217
In [157]: grouped_trans.count() # counts after transformation
Out[157]:
A B C
GR 228 228 228
JP 267 267 267
UK 247 247 247
US 258 258 258
In [158]: grouped_trans.size() # Verify non-NA count equals group size
Out[158]:
GR 228
JP 267
UK 247
US 258
dtype: int64
As mentioned in the note above, each of the examples in this section can be computed more efficiently using built-in methods. In the code below, the inefficient way using a UDF is commented out and the faster alternative appears below.
# ts.groupby(lambda x: x.year).transform(
# lambda x: (x - x.mean()) / x.std()
# )
In [159]: grouped = ts.groupby(lambda x: x.year)
In [160]: result = (ts - grouped.transform("mean")) / grouped.transform("std")
# ts.groupby(lambda x: x.year).transform(lambda x: x.max() - x.min())
In [161]: grouped = ts.groupby(lambda x: x.year)
In [162]: result = grouped.transform("max") - grouped.transform("min")
# grouped = data_df.groupby(key)
# grouped.transform(lambda x: x.fillna(x.mean()))
In [163]: grouped = data_df.groupby(key)
In [164]: result = data_df.fillna(grouped.transform("mean"))
Window and resample operations#
It is possible to use resample(), expanding() and
rolling() as methods on groupbys.
The example below will apply the rolling() method on the samples of
the column B, based on the groups of column A.
In [165]: df_re = pd.DataFrame({"A": [1] * 10 + [5] * 10, "B": np.arange(20)})
In [166]: df_re
Out[166]:
A B
0 1 0
1 1 1
2 1 2
3 1 3
4 1 4
.. .. ..
15 5 15
16 5 16
17 5 17
18 5 18
19 5 19
[20 rows x 2 columns]
In [167]: df_re.groupby("A").rolling(4).B.mean()
Out[167]:
A
1 0 NaN
1 NaN
2 NaN
3 1.5
4 2.5
...
5 15 13.5
16 14.5
17 15.5
18 16.5
19 17.5
Name: B, Length: 20, dtype: float64
The expanding() method will accumulate a given operation
(sum() in the example) for all the members of each particular
group.
In [168]: df_re.groupby("A").expanding().sum()
Out[168]:
B
A
1 0 0.0
1 1.0
2 3.0
3 6.0
4 10.0
... ...
5 15 75.0
16 91.0
17 108.0
18 126.0
19 145.0
[20 rows x 1 columns]
Suppose you want to use the resample() method to get a daily
frequency in each group of your dataframe, and wish to complete the
missing values with the ffill() method.
In [169]: df_re = pd.DataFrame(
.....: {
.....: "date": pd.date_range(start="2016-01-01", periods=4, freq="W"),
.....: "group": [1, 1, 2, 2],
.....: "val": [5, 6, 7, 8],
.....: }
.....: ).set_index("date")
.....:
In [170]: df_re
Out[170]:
group val
date
2016-01-03 1 5
2016-01-10 1 6
2016-01-17 2 7
2016-01-24 2 8
In [171]: df_re.groupby("group").resample("1D").ffill()
Out[171]:
group val
group date
1 2016-01-03 1 5
2016-01-04 1 5
2016-01-05 1 5
2016-01-06 1 5
2016-01-07 1 5
... ... ...
2 2016-01-20 2 7
2016-01-21 2 7
2016-01-22 2 7
2016-01-23 2 7
2016-01-24 2 8
[16 rows x 2 columns]
Filtration#
A filtration is a GroupBy operation the subsets the original grouping object. It
may either filter out entire groups, part of groups, or both. Filtrations return
a filtered version of the calling object, including the grouping columns when provided.
In the following example, class is included in the result.
In [172]: speeds
Out[172]:
class order max_speed
falcon bird Falconiformes 389.0
parrot bird Psittaciformes 24.0
lion mammal Carnivora 80.2
monkey mammal Primates NaN
leopard mammal Carnivora 58.0
In [173]: speeds.groupby("class").nth(1)
Out[173]:
class order max_speed
parrot bird Psittaciformes 24.0
monkey mammal Primates NaN
Note
Unlike aggregations, filtrations do not add the group keys to the index of the
result. Because of this, passing as_index=False or sort=True will not
affect these methods.
Filtrations will respect subsetting the columns of the GroupBy object.
In [174]: speeds.groupby("class")[["order", "max_speed"]].nth(1)
Out[174]:
order max_speed
parrot Psittaciformes 24.0
monkey Primates NaN
Built-in filtrations#
The following methods on GroupBy act as filtrations. All these methods have a Cython-optimized implementation.
Method |
Description |
|---|---|
Select the top row(s) of each group |
|
Select the nth row(s) of each group |
|
Select the bottom row(s) of each group |
Users can also use transformations along with Boolean indexing to construct complex filtrations within groups. For example, suppose we are given groups of products and their volumes, and we wish to subset the data to only the largest products capturing no more than 90% of the total volume within each group.
In [175]: product_volumes = pd.DataFrame(
.....: {
.....: "group": list("xxxxyyy"),
.....: "product": list("abcdefg"),
.....: "volume": [10, 30, 20, 15, 40, 10, 20],
.....: }
.....: )
.....:
In [176]: product_volumes
Out[176]:
group product volume
0 x a 10
1 x b 30
2 x c 20
3 x d 15
4 y e 40
5 y f 10
6 y g 20
# Sort by volume to select the largest products first
In [177]: product_volumes = product_volumes.sort_values("volume", ascending=False)
In [178]: grouped = product_volumes.groupby("group")["volume"]
In [179]: cumpct = grouped.cumsum() / grouped.transform("sum")
In [180]: cumpct
Out[180]:
4 0.571429
1 0.400000
2 0.666667
6 0.857143
3 0.866667
0 1.000000
5 1.000000
Name: volume, dtype: float64
In [181]: significant_products = product_volumes[cumpct <= 0.9]
In [182]: significant_products.sort_values(["group", "product"])
Out[182]:
group product volume
1 x b 30
2 x c 20
3 x d 15
4 y e 40
6 y g 20
The filter method#
Note
Filtering by supplying filter with a User-Defined Function (UDF) is
often less performant than using the built-in methods on GroupBy.
Consider breaking up a complex operation into a chain of operations that utilize
the built-in methods.
The filter method takes a User-Defined Function (UDF) that, when applied to
an entire group, returns either True or False. The result of the filter
method is then the subset of groups for which the UDF returned True.
Suppose we want to take only elements that belong to groups with a group sum greater than 2.
In [183]: sf = pd.Series([1, 1, 2, 3, 3, 3])
In [184]: sf.groupby(sf).filter(lambda x: x.sum() > 2)
Out[184]:
3 3
4 3
5 3
dtype: int64
Another useful operation is filtering out elements that belong to groups with only a couple members.
In [185]: dff = pd.DataFrame({"A": np.arange(8), "B": list("aabbbbcc")})
In [186]: dff.groupby("B").filter(lambda x: len(x) > 2)
Out[186]:
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 [187]: dff.groupby("B").filter(lambda x: len(x) > 2, dropna=False)
Out[187]:
A B
0 NaN NaN
1 NaN NaN
2 2.0 b
3 3.0 b
4 4.0 b
5 5.0 b
6 NaN NaN
7 NaN NaN
For DataFrames with multiple columns, filters should explicitly specify a column as the filter criterion.
In [188]: dff["C"] = np.arange(8)
In [189]: dff.groupby("B").filter(lambda x: len(x["C"]) > 2)
Out[189]:
A B C
2 2 b 2
3 3 b 3
4 4 b 4
5 5 b 5
Flexible apply#
Some operations on the grouped data might not fit into the aggregation,
transformation, or filtration categories. For these, you can use the apply
function.
Warning
apply has to try to infer from the result whether it should act as a reducer,
transformer, or filter, depending on exactly what is passed to it. Thus the
grouped column(s) may be included in the output or not. While
it tries to intelligently guess how to behave, it can sometimes guess wrong.
Note
All of the examples in this section can be more reliably, and more efficiently, computed using other pandas functionality.
In [190]: df
Out[190]:
A B C D
0 foo one -0.575247 1.346061
1 bar one 0.254161 1.511763
2 foo two -1.143704 1.627081
3 bar three 0.215897 -0.990582
4 foo two 1.193555 -0.441652
5 bar two -0.077118 1.211526
6 foo one -0.408530 0.268520
7 foo three -0.862495 0.024580
In [191]: grouped = df.groupby("A")
# could also just call .describe()
In [192]: grouped["C"].apply(lambda x: x.describe())
Out[192]:
A
bar count 3.000000
mean 0.130980
std 0.181231
min -0.077118
25% 0.069390
...
foo min -1.143704
25% -0.862495
50% -0.575247
75% -0.408530
max 1.193555
Name: C, Length: 16, dtype: float64
The dimension of the returned result can also change:
In [193]: grouped = df.groupby('A')['C']
In [194]: def f(group):
.....: return pd.DataFrame({'original': group,
.....: 'demeaned': group - group.mean()})
.....:
In [195]: grouped.apply(f)
Out[195]:
original demeaned
A
bar 1 0.254161 0.123181
3 0.215897 0.084917
5 -0.077118 -0.208098
foo 0 -0.575247 -0.215962
2 -1.143704 -0.784420
4 1.193555 1.552839
6 -0.408530 -0.049245
7 -0.862495 -0.503211
Similar to The aggregate() method, the resulting dtype will reflect that of the
apply function. If the results from different groups have different dtypes, then
a common dtype will be determined in the same way as DataFrame construction.
Control grouped column(s) placement with group_keys#
To control whether the grouped column(s) are included in the indices, you can use
the argument group_keys which defaults to True. Compare
In [196]: df.groupby("A", group_keys=True).apply(lambda x: x)
Out[196]:
A B C D
A
bar 1 bar one 0.254161 1.511763
3 bar three 0.215897 -0.990582
5 bar two -0.077118 1.211526
foo 0 foo one -0.575247 1.346061
2 foo two -1.143704 1.627081
4 foo two 1.193555 -0.441652
6 foo one -0.408530 0.268520
7 foo three -0.862495 0.024580
with
In [197]: df.groupby("A", group_keys=False).apply(lambda x: x)
Out[197]:
A B C D
0 foo one -0.575247 1.346061
1 bar one 0.254161 1.511763
2 foo two -1.143704 1.627081
3 bar three 0.215897 -0.990582
4 foo two 1.193555 -0.441652
5 bar two -0.077118 1.211526
6 foo one -0.408530 0.268520
7 foo three -0.862495 0.024580
Numba Accelerated Routines#
New in version 1.1.
If Numba is installed as an optional dependency, the transform and
aggregate methods support engine='numba' and engine_kwargs arguments.
See enhancing performance with Numba for general usage of the arguments
and performance considerations.
The function signature must start with values, index exactly as the data belonging to each group
will be passed into values, and the group index will be passed into index.
Warning
When using engine='numba', there will be no “fall back” behavior internally. The group
data and group index will be passed as NumPy arrays to the JITed user defined function, and no
alternative execution attempts will be tried.
Other useful features#
Exclusion of “nuisance” columns#
Again consider the example DataFrame we’ve been looking at:
In [198]: df
Out[198]:
A B C D
0 foo one -0.575247 1.346061
1 bar one 0.254161 1.511763
2 foo two -1.143704 1.627081
3 bar three 0.215897 -0.990582
4 foo two 1.193555 -0.441652
5 bar two -0.077118 1.211526
6 foo one -0.408530 0.268520
7 foo three -0.862495 0.024580
Suppose we wish 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 because it is not numeric. We refer to these non-numeric columns as
“nuisance” columns. You can avoid nuisance columns by specifying numeric_only=True:
In [199]: df.groupby("A").std(numeric_only=True)
Out[199]:
C D
A
bar 0.181231 1.366330
foo 0.912265 0.884785
Note that df.groupby('A').colname.std(). is more efficient than
df.groupby('A').std().colname. So if the result of an aggregation function
is only needed over one column (here colname), it may be filtered
before applying the aggregation function.
In [200]: from decimal import Decimal
In [201]: df_dec = pd.DataFrame(
.....: {
.....: "id": [1, 2, 1, 2],
.....: "int_column": [1, 2, 3, 4],
.....: "dec_column": [
.....: Decimal("0.50"),
.....: Decimal("0.15"),
.....: Decimal("0.25"),
.....: Decimal("0.40"),
.....: ],
.....: }
.....: )
.....:
# Decimal columns can be sum'd explicitly by themselves...
In [202]: df_dec.groupby(["id"])[["dec_column"]].sum()
Out[202]:
dec_column
id
1 0.75
2 0.55
# ...but cannot be combined with standard data types or they will be excluded
In [203]: df_dec.groupby(["id"])[["int_column", "dec_column"]].sum()
Out[203]:
int_column dec_column
id
1 4 0.75
2 6 0.55
# Use .agg function to aggregate over standard and "nuisance" data types
# at the same time
In [204]: df_dec.groupby(["id"]).agg({"int_column": "sum", "dec_column": "sum"})
Out[204]:
int_column dec_column
id
1 4 0.75
2 6 0.55
Handling of (un)observed Categorical values#
When using a Categorical grouper (as a single grouper, or as part of multiple groupers), the observed keyword
controls whether to return a cartesian product of all possible groupers values (observed=False) or only those
that are observed groupers (observed=True).
Show all values:
In [205]: pd.Series([1, 1, 1]).groupby(
.....: pd.Categorical(["a", "a", "a"], categories=["a", "b"]), observed=False
.....: ).count()
.....:
Out[205]:
a 3
b 0
dtype: int64
Show only the observed values:
In [206]: pd.Series([1, 1, 1]).groupby(
.....: pd.Categorical(["a", "a", "a"], categories=["a", "b"]), observed=True
.....: ).count()
.....:
Out[206]:
a 3
dtype: int64
The returned dtype of the grouped will always include all of the categories that were grouped.
In [207]: s = (
.....: pd.Series([1, 1, 1])
.....: .groupby(pd.Categorical(["a", "a", "a"], categories=["a", "b"]), observed=False)
.....: .count()
.....: )
.....:
In [208]: s.index.dtype
Out[208]: CategoricalDtype(categories=['a', 'b'], ordered=False, categories_dtype=object)
NA and NaT group handling#
If there are any NaN or NaT values in the grouping key, these will be automatically excluded. In other words, there will never be an “NA group” or “NaT 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 instances of pandas’s Categorical class
can be used as group keys. If so, the order of the levels will be preserved:
In [209]: data = pd.Series(np.random.randn(100))
In [210]: factor = pd.qcut(data, [0, 0.25, 0.5, 0.75, 1.0])
In [211]: data.groupby(factor, observed=False).mean()
Out[211]:
(-2.131, -0.461] -1.049409
(-0.461, 0.0721] -0.125522
(0.0721, 0.739] 0.404226
(0.739, 2.022] 1.202518
dtype: float64
Grouping with a grouper specification#
You may need to specify a bit more data to properly group. You can
use the pd.Grouper to provide this local control.
In [212]: import datetime
In [213]: df = pd.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": [
.....: datetime.datetime(2013, 1, 1, 13, 0),
.....: datetime.datetime(2013, 1, 1, 13, 5),
.....: datetime.datetime(2013, 10, 1, 20, 0),
.....: datetime.datetime(2013, 10, 2, 10, 0),
.....: datetime.datetime(2013, 10, 1, 20, 0),
.....: datetime.datetime(2013, 10, 2, 10, 0),
.....: datetime.datetime(2013, 12, 2, 12, 0),
.....: datetime.datetime(2013, 12, 2, 14, 0),
.....: ],
.....: }
.....: )
.....:
In [214]: df
Out[214]:
Branch Buyer Quantity Date
0 A Carl 1 2013-01-01 13:00:00
1 A Mark 3 2013-01-01 13:05:00
2 A Carl 5 2013-10-01 20:00:00
3 A Carl 1 2013-10-02 10:00:00
4 A Joe 8 2013-10-01 20:00:00
5 A Joe 1 2013-10-02 10:00:00
6 A Joe 9 2013-12-02 12:00:00
7 B Carl 3 2013-12-02 14:00:00
Groupby a specific column with the desired frequency. This is like resampling.
In [215]: df.groupby([pd.Grouper(freq="1M", key="Date"), "Buyer"])[["Quantity"]].sum()
Out[215]:
Quantity
Date Buyer
2013-01-31 Carl 1
Mark 3
2013-10-31 Carl 6
Joe 9
2013-12-31 Carl 3
Joe 9
When freq is specified, the object returned by pd.Grouper will be an
instance of pandas.api.typing.TimeGrouper. You have an ambiguous specification
in that you have a named index and a column that could be potential groupers.
In [216]: df = df.set_index("Date")
In [217]: df["Date"] = df.index + pd.offsets.MonthEnd(2)
In [218]: df.groupby([pd.Grouper(freq="6M", key="Date"), "Buyer"])[["Quantity"]].sum()
Out[218]:
Quantity
Date Buyer
2013-02-28 Carl 1
Mark 3
2014-02-28 Carl 9
Joe 18
In [219]: df.groupby([pd.Grouper(freq="6M", level="Date"), "Buyer"])[["Quantity"]].sum()
Out[219]:
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 [220]: df = pd.DataFrame([[1, 2], [1, 4], [5, 6]], columns=["A", "B"])
In [221]: df
Out[221]:
A B
0 1 2
1 1 4
2 5 6
In [222]: g = df.groupby("A")
In [223]: g.head(1)
Out[223]:
A B
0 1 2
2 5 6
In [224]: g.tail(1)
Out[224]:
A B
1 1 4
2 5 6
This shows the first or last n rows from each group.
Taking the nth row of each group#
To select the nth item from each group, use DataFrameGroupBy.nth() or
SeriesGroupBy.nth(). Arguments supplied can be any integer, lists of integers,
slices, or lists of slices; see below for examples. When the nth element of a group
does not exist an error is not raised; instead no corresponding rows are returned.
In general this operation acts as a filtration. In certain cases it will also return one row per group, making it also a reduction. However because in general it can return zero or multiple rows per group, pandas treats it as a filtration in all cases.
In [225]: df = pd.DataFrame([[1, np.nan], [1, 4], [5, 6]], columns=["A", "B"])
In [226]: g = df.groupby("A")
In [227]: g.nth(0)
Out[227]:
A B
0 1 NaN
2 5 6.0
In [228]: g.nth(-1)
Out[228]:
A B
1 1 4.0
2 5 6.0
In [229]: g.nth(1)
Out[229]:
A B
1 1 4.0
If the nth element of a group does not exist, then no corresponding row is included
in the result. In particular, if the specified n is larger than any group, the
result will be an empty DataFrame.
In [230]: g.nth(5)
Out[230]:
Empty DataFrame
Columns: [A, B]
Index: []
If you want to select the nth not-null item, use the dropna kwarg. For a DataFrame this should be either 'any' or 'all' just like you would pass to dropna:
# nth(0) is the same as g.first()
In [231]: g.nth(0, dropna="any")
Out[231]:
A B
1 1 4.0
2 5 6.0
In [232]: g.first()
Out[232]:
B
A
1 4.0
5 6.0
# nth(-1) is the same as g.last()
In [233]: g.nth(-1, dropna="any")
Out[233]:
A B
1 1 4.0
2 5 6.0
In [234]: g.last()
Out[234]:
B
A
1 4.0
5 6.0
In [235]: g.B.nth(0, dropna="all")
Out[235]:
1 4.0
2 6.0
Name: B, dtype: float64
You can also select multiple rows from each group by specifying multiple nth values as a list of ints.
In [236]: business_dates = pd.date_range(start="4/1/2014", end="6/30/2014", freq="B")
In [237]: df = pd.DataFrame(1, index=business_dates, columns=["a", "b"])
# get the first, 4th, and last date index for each month
In [238]: df.groupby([df.index.year, df.index.month]).nth([0, 3, -1])
Out[238]:
a b
2014-04-01 1 1
2014-04-04 1 1
2014-04-30 1 1
2014-05-01 1 1
2014-05-06 1 1
2014-05-30 1 1
2014-06-02 1 1
2014-06-05 1 1
2014-06-30 1 1
You may also use slices or lists of slices.
In [239]: df.groupby([df.index.year, df.index.month]).nth[1:]
Out[239]:
a b
2014-04-02 1 1
2014-04-03 1 1
2014-04-04 1 1
2014-04-07 1 1
2014-04-08 1 1
... .. ..
2014-06-24 1 1
2014-06-25 1 1
2014-06-26 1 1
2014-06-27 1 1
2014-06-30 1 1
[62 rows x 2 columns]
In [240]: df.groupby([df.index.year, df.index.month]).nth[1:, :-1]
Out[240]:
a b
2014-04-01 1 1
2014-04-02 1 1
2014-04-03 1 1
2014-04-04 1 1
2014-04-07 1 1
... .. ..
2014-06-24 1 1
2014-06-25 1 1
2014-06-26 1 1
2014-06-27 1 1
2014-06-30 1 1
[65 rows x 2 columns]
Enumerate group items#
To see the order in which each row appears within its group, use the
cumcount method:
In [241]: dfg = pd.DataFrame(list("aaabba"), columns=["A"])
In [242]: dfg
Out[242]:
A
0 a
1 a
2 a
3 b
4 b
5 a
In [243]: dfg.groupby("A").cumcount()
Out[243]:
0 0
1 1
2 2
3 0
4 1
5 3
dtype: int64
In [244]: dfg.groupby("A").cumcount(ascending=False)
Out[244]:
0 3
1 2
2 1
3 1
4 0
5 0
dtype: int64
Enumerate groups#
To see the ordering of the groups (as opposed to the order of rows
within a group given by cumcount) you can use
ngroup().
Note that the numbers given to the groups match the order in which the groups would be seen when iterating over the groupby object, not the order they are first observed.
In [245]: dfg = pd.DataFrame(list("aaabba"), columns=["A"])
In [246]: dfg
Out[246]:
A
0 a
1 a
2 a
3 b
4 b
5 a
In [247]: dfg.groupby("A").ngroup()
Out[247]:
0 0
1 0
2 0
3 1
4 1
5 0
dtype: int64
In [248]: dfg.groupby("A").ngroup(ascending=False)
Out[248]:
0 1
1 1
2 1
3 0
4 0
5 1
dtype: int64
Plotting#
Groupby also works with some plotting methods. In this case, suppose we suspect that the values in column 1 are 3 times higher on average in group “B”.
In [249]: np.random.seed(1234)
In [250]: df = pd.DataFrame(np.random.randn(50, 2))
In [251]: df["g"] = np.random.choice(["A", "B"], size=50)
In [252]: df.loc[df["g"] == "B", 1] += 3
We can easily visualize this with a boxplot:
In [253]: df.groupby("g").boxplot()
Out[253]:
A Axes(0.1,0.15;0.363636x0.75)
B Axes(0.536364,0.15;0.363636x0.75)
dtype: object
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.
Piping function calls#
Similar to the functionality provided by DataFrame and Series, functions
that take GroupBy objects can be chained together using a pipe method to
allow for a cleaner, more readable syntax. To read about .pipe in general terms,
see here.
Combining .groupby and .pipe is often useful when you need to reuse
GroupBy objects.
As an example, imagine having a DataFrame with columns for stores, products, revenue and quantity sold. We’d like to do a groupwise calculation of prices (i.e. revenue/quantity) per store and per product. We could do this in a multi-step operation, but expressing it in terms of piping can make the code more readable. First we set the data:
In [254]: n = 1000
In [255]: df = pd.DataFrame(
.....: {
.....: "Store": np.random.choice(["Store_1", "Store_2"], n),
.....: "Product": np.random.choice(["Product_1", "Product_2"], n),
.....: "Revenue": (np.random.random(n) * 50 + 10).round(2),
.....: "Quantity": np.random.randint(1, 10, size=n),
.....: }
.....: )
.....:
In [256]: df.head(2)
Out[256]:
Store Product Revenue Quantity
0 Store_2 Product_1 26.12 1
1 Store_2 Product_1 28.86 1
Now, to find prices per store/product, we can simply do:
In [257]: (
.....: df.groupby(["Store", "Product"])
.....: .pipe(lambda grp: grp.Revenue.sum() / grp.Quantity.sum())
.....: .unstack()
.....: .round(2)
.....: )
.....:
Out[257]:
Product Product_1 Product_2
Store
Store_1 6.82 7.05
Store_2 6.30 6.64
Piping can also be expressive when you want to deliver a grouped object to some arbitrary function, for example:
In [258]: def mean(groupby):
.....: return groupby.mean()
.....:
In [259]: df.groupby(["Store", "Product"]).pipe(mean)
Out[259]:
Revenue Quantity
Store Product
Store_1 Product_1 34.622727 5.075758
Product_2 35.482815 5.029630
Store_2 Product_1 32.972837 5.237589
Product_2 34.684360 5.224000
Here mean takes a GroupBy object and finds the mean of the Revenue and Quantity
columns respectively for each Store-Product combination. The mean function can
be any function that takes in a GroupBy object; the .pipe will pass the GroupBy
object as a parameter into the function you specify.
Examples#
Regrouping by factor#
Regroup columns of a DataFrame according to their sum, and sum the aggregated ones.
In [260]: df = pd.DataFrame({"a": [1, 0, 0], "b": [0, 1, 0], "c": [1, 0, 0], "d": [2, 3, 4]})
In [261]: df
Out[261]:
a b c d
0 1 0 1 2
1 0 1 0 3
2 0 0 0 4
In [262]: dft = df.T
In [263]: dft.groupby(dft.sum()).sum()
Out[263]:
Empty DataFrame
Columns: [0, 1, 2]
Index: []
Multi-column factorization#
By using ngroup(), we can extract
information about the groups in a way similar to factorize() (as described
further in the reshaping API) but which applies
naturally to multiple columns of mixed type and different
sources. This can be useful as an intermediate categorical-like step
in processing, when the relationships between the group rows are more
important than their content, or as input to an algorithm which only
accepts the integer encoding. (For more information about support in
pandas for full categorical data, see the Categorical
introduction and the
API documentation.)
In [264]: dfg = pd.DataFrame({"A": [1, 1, 2, 3, 2], "B": list("aaaba")})
In [265]: dfg
Out[265]:
A B
0 1 a
1 1 a
2 2 a
3 3 b
4 2 a
In [266]: dfg.groupby(["A", "B"]).ngroup()
Out[266]:
0 0
1 0
2 1
3 2
4 1
dtype: int64
In [267]: dfg.groupby(["A", [0, 0, 0, 1, 1]]).ngroup()
Out[267]:
0 0
1 0
2 1
3 3
4 2
dtype: int64
Groupby by indexer to ‘resample’ data#
Resampling produces new hypothetical samples (resamples) from already existing observed data or from a model that generates data. These new samples are similar to the pre-existing samples.
In order for resample to work on indices that are non-datetimelike, the following procedure can be utilized.
In the following examples, df.index // 5 returns a binary array which is used to determine what gets selected for the groupby operation.
Note
The example below shows how we can downsample by consolidation of samples into fewer ones. Here by using df.index // 5, we are aggregating the samples in bins. By applying std() function, we aggregate the information contained in many samples into a small subset of values which is their standard deviation thereby reducing the number of samples.
In [268]: df = pd.DataFrame(np.random.randn(10, 2))
In [269]: df
Out[269]:
0 1
0 -0.793893 0.321153
1 0.342250 1.618906
2 -0.975807 1.918201
3 -0.810847 -1.405919
4 -1.977759 0.461659
5 0.730057 -1.316938
6 -0.751328 0.528290
7 -0.257759 -1.081009
8 0.505895 -1.701948
9 -1.006349 0.020208
In [270]: df.index // 5
Out[270]: Index([0, 0, 0, 0, 0, 1, 1, 1, 1, 1], dtype='int64')
In [271]: df.groupby(df.index // 5).std()
Out[271]:
0 1
0 0.823647 1.312912
1 0.760109 0.942941
Returning a Series to propagate 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 [272]: 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 [273]: def compute_metrics(x):
.....: result = {"b_sum": x["b"].sum(), "c_mean": x["c"].mean()}
.....: return pd.Series(result, name="metrics")
.....:
In [274]: result = df.groupby("a").apply(compute_metrics)
In [275]: result
Out[275]:
metrics b_sum c_mean
a
0 2.0 0.5
1 2.0 0.5
2 2.0 0.5
In [276]: result.stack(future_stack=True)
Out[276]:
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