Pandas分组与聚合
分组 (groupby)
对数据集进行分组，然后对每组进行统计分析
SQL能够对数据进行过滤，分组聚合
pandas能利用groupby进行更加复杂的分组运算
分组运算过程：split->apply->combine
拆分：进行分组的根据
应用：每个分组运行的计算规则
合并：把每个分组的计算结果合并起来
示例代码：
import pandas as pd
import numpy as np
dict_obj = {'key1' : ['a', 'b', 'a', 'b',
'a', 'b', 'a', 'a'],
'key2' : ['one', 'one', 'two', 'three',
'two', 'two', 'one', 'three'],
'data1': np.random.randn(8),
'data2': np.random.randn(8)}
df_obj = pd.DataFrame(dict_obj)
print(df_obj)
运行结果：
data1 data2 key1 key2
0 0.974685 -0.672494 a one
1 -0.214324 0.758372 b one
2 1.508838 0.392787 a two
3 0.522911 0.630814 b three
4 1.347359 -0.177858 a two
5 -0.264616 1.017155 b two
6 -0.624708 0.450885 a one
7 -1.019229 -1.143825 a three
一、GroupBy对象：DataFrameGroupBy，SeriesGroupBy
1. 分组操作
groupby()进行分组，GroupBy对象没有进行实际运算，只是包含分组的中间数据
按列名分组：obj.groupby(‘label’)
示例代码：
# dataframe根据key1进行分组
print(type(df_obj.groupby('key1')))
# dataframe的 data1 列根据 key1 进行分组
print(type(df_obj['data1'].groupby(df_obj['key1'])))
运行结果：
<class 'pandas.core.groupby.DataFrameGroupBy'>
<class 'pandas.core.groupby.SeriesGroupBy'>
2. 分组运算
对GroupBy对象进行分组运算/多重分组运算，如mean()
非数值数据不进行分组运算
示例代码：
# 分组运算
grouped1 = df_obj.groupby('key1')
print(grouped1.mean())
grouped2 = df_obj['data1'].groupby(df_obj['key1'])
print(grouped2.mean())
运行结果：
data1 data2
key1
a 0.437389 -0.230101
b 0.014657 0.802114
key1
a 0.437389
b 0.014657
Name: data1, dtype: float64
size() 返回每个分组的元素个数
示例代码：
# size
print(grouped1.size())
print(grouped2.size())
运行结果：
key1
a 5
b 3
dtype: int64
key1
a 5
b 3
dtype: int64
3. 按自定义的key分组
obj.groupby(self_def_key)
自定义的key可为列表或多层列表
obj.groupby([‘label1’, ‘label2’])->多层dataframe
示例代码：
# 按自定义key分组，列表
self_def_key = [0, 1, 2, 3, 3, 4, 5, 7]
print(df_obj.groupby(self_def_key).size())
# 按自定义key分组，多层列表
print(df_obj.groupby([df_obj['key1'], df_obj['key2']]).size())
# 按多个列多层分组
grouped2 = df_obj.groupby(['key1', 'key2'])
print(grouped2.size())
# 多层分组按key的顺序进行
grouped3 = df_obj.groupby(['key2', 'key1'])
print(grouped3.mean())
# unstack可以将多层索引的结果转换成单层的dataframe
print(grouped3.mean().unstack())
运行结果：
0 1
1 1
2 1
3 2
4 1
5 1
7 1
dtype: int64
key1 key2
a one 2
three 1
two 2
b one 1
three 1
two 1
dtype: int64
key1 key2
a one 2
three 1
two 2
b one 1
three 1
two 1
dtype: int64
data1 data2
key2 key1
one a 0.174988 -0.110804
b -0.214324 0.758372
three a -1.019229 -1.143825
b 0.522911 0.630814
two a 1.428099 0.107465
b -0.264616 1.017155
data1 data2
key1 a b a b
key2
one 0.174988 -0.214324 -0.110804 0.758372
three -1.019229 0.522911 -1.143825 0.630814
two 1.428099 -0.264616 0.107465 1.017155
二、GroupBy对象支持迭代操作
每次迭代返回一个元组 (group_name, group_data)
可用于分组数据的具体运算
1. 单层分组
示例代码：
# 单层分组，根据key1
for group_name, group_data in grouped1:
print(group_name)
print(group_data)
运行结果：
a
data1 data2 key1 key2
0 0.974685 -0.672494 a one
2 1.508838 0.392787 a two
4 1.347359 -0.177858 a two
6 -0.624708 0.450885 a one
7 -1.019229 -1.143825 a three
b
data1 data2 key1 key2
1 -0.214324 0.758372 b one
3 0.522911 0.630814 b three
5 -0.264616 1.017155 b two
2. 多层分组
示例代码：
# 多层分组，根据key1 和 key2
for group_name, group_data in grouped2:
print(group_name)
print(group_data)
运行结果：
('a', 'one')
data1 data2 key1 key2
0 0.974685 -0.672494 a one
6 -0.624708 0.450885 a one
('a', 'three')
data1 data2 key1 key2
7 -1.019229 -1.143825 a three
('a', 'two')
data1 data2 key1 key2
2 1.508838 0.392787 a two
4 1.347359 -0.177858 a two
('b', 'one')
data1 data2 key1 key2
1 -0.214324 0.758372 b one
('b', 'three')
data1 data2 key1 key2
3 0.522911 0.630814 b three
('b', 'two')
data1 data2 key1 key2
5 -0.264616 1.017155 b two
三、GroupBy对象可以转换成列表或字典
示例代码：
# GroupBy对象转换list
print(list(grouped1))
# GroupBy对象转换dict
print(dict(list(grouped1)))
运行结果：
[('a', data1 data2 key1 key2
0 0.974685 -0.672494 a one
2 1.508838 0.392787 a two
4 1.347359 -0.177858 a two
6 -0.624708 0.450885 a one
7 -1.019229 -1.143825 a three),
('b', data1 data2 key1 key2
1 -0.214324 0.758372 b one
3 0.522911 0.630814 b three
5 -0.264616 1.017155 b two)]
{'a': data1 data2 key1 key2
0 0.974685 -0.672494 a one
2 1.508838 0.392787 a two
4 1.347359 -0.177858 a two
6 -0.624708 0.450885 a one
7 -1.019229 -1.143825 a three,
'b': data1 data2 key1 key2
1 -0.214324 0.758372 b one
3 0.522911 0.630814 b three
5 -0.264616 1.017155 b two}
1. 按列分组、按数据类型分组
示例代码：
# 按列分组
print(df_obj.dtypes)
# 按数据类型分组
print(df_obj.groupby(df_obj.dtypes, axis=1).size())
print(df_obj.groupby(df_obj.dtypes, axis=1).sum())
运行结果：
data1 float64
data2 float64
key1 object
key2 object
dtype: object
float64 2
object 2
dtype: int64
float64 object
0 0.302191 a one
1 0.544048 b one
2 1.901626 a two
3 1.153725 b three
4 1.169501 a two
5 0.752539 b two
6 -0.173823 a one
7 -2.163054 a three
2. 其他分组方法
示例代码：
df_obj2 = pd.DataFrame(np.random.randint(1, 10, (5,5)),
columns=['a', 'b', 'c', 'd', 'e'],
index=['A', 'B', 'C', 'D', 'E'])
df_obj2.ix[1, 1:4] = np.NaN
print(df_obj2)
运行结果：
a b c d e
A 7 2.0 4.0 5.0 8
B 4 NaN NaN NaN 1
C 3 2.0 5.0 4.0 6
D 3 1.0 9.0 7.0 3
E 6 1.0 6.0 8.0 1
3. 通过字典分组
示例代码：
# 通过字典分组
mapping_dict = {'a':'Python', 'b':'Python', 'c':'Java', 'd':'C', 'e':'Java'}
print(df_obj2.groupby(mapping_dict, axis=1).size())
print(df_obj2.groupby(mapping_dict, axis=1).count()) # 非NaN的个数
print(df_obj2.groupby(mapping_dict, axis=1).sum())
运行结果：
C 1
Java 2
Python 2
dtype: int64
C Java Python
A 1 2 2
B 0 1 1
C 1 2 2
D 1 2 2
E 1 2 2
C Java Python
A 5.0 12.0 9.0
B NaN 1.0 4.0
C 4.0 11.0 5.0
D 7.0 12.0 4.0
E 8.0 7.0 7.0
4. 通过函数分组，函数传入的参数为行索引或列索引
示例代码：
# 通过函数分组
df_obj3 = pd.DataFrame(np.random.randint(1, 10, (5,5)),
columns=['a', 'b', 'c', 'd', 'e'],
index=['AA', 'BBB', 'CC', 'D', 'EE'])
#df_obj3
def group_key(idx):
"""
idx 为列索引或行索引
"""
#return idx
return len(idx)
print(df_obj3.groupby(group_key).size())
# 以上自定义函数等价于
#df_obj3.groupby(len).size()
运行结果：
1 1
2 3
3 1
dtype: int64
5. 通过索引级别分组
示例代码：
# 通过索引级别分组
columns = pd.MultiIndex.from_arrays([['Python', 'Java', 'Python', 'Java', 'Python'],
['A', 'A', 'B', 'C', 'B']], names=['language', 'index'])
df_obj4 = pd.DataFrame(np.random.randint(1, 10, (5, 5)), columns=columns)
print(df_obj4)
# 根据language进行分组
print(df_obj4.groupby(level='language', axis=1).sum())
# 根据index进行分组
print(df_obj4.groupby(level='index', axis=1).sum())
运行结果：
language Python Java Python Java Python
index A A B C B
0 2 7 8 4 3
1 5 2 6 1 2
2 6 4 4 5 2
3 4 7 4 3 1
4 7 4 3 4 8
language Java Python
0 11 13
1 3 13
2 9 12
3 10 9
4 8 18
index A B C
0 9 11 4
1 7 8 1
2 10 6 5
3 11 5 3
4 11 11 4
聚合 (aggregation)
数组产生标量的过程，如mean()、count()等
常用于对分组后的数据进行计算
示例代码：
dict_obj = {'key1' : ['a', 'b', 'a', 'b',
'a', 'b', 'a', 'a'],
'key2' : ['one', 'one', 'two', 'three',
'two', 'two', 'one', 'three'],
'data1': np.random.randint(1,10, 8),
'data2': np.random.randint(1,10, 8)}
df_obj5 = pd.DataFrame(dict_obj)
print(df_obj5)
运行结果：
data1 data2 key1 key2
0 3 7 a one
1 1 5 b one
2 7 4 a two
3 2 4 b three
4 6 4 a two
5 9 9 b two
6 3 5 a one
7 8 4 a three
1. 内置的聚合函数
sum(), mean(), max(), min(), count(), size(), describe()
示例代码：
print(df_obj5.groupby('key1').sum())
print(df_obj5.groupby('key1').max())
print(df_obj5.groupby('key1').min())
print(df_obj5.groupby('key1').mean())
print(df_obj5.groupby('key1').size())
print(df_obj5.groupby('key1').count())
print(df_obj5.groupby('key1').describe())
运行结果：
data1 data2
key1
a 27 24
b 12 18
data1 data2 key2
key1
a 8 7 two
b 9 9 two
data1 data2 key2
key1
a 3 4 one
b 1 4 one
data1 data2
key1
a 5.4 4.8
b 4.0 6.0
key1
a 5
b 3
dtype: int64
data1 data2 key2
key1
a 5 5 5
b 3 3 3
data1 data2
key1
a count 5.000000 5.000000
mean 5.400000 4.800000
std 2.302173 1.303840
min 3.000000 4.000000
25% 3.000000 4.000000
50% 6.000000 4.000000
75% 7.000000 5.000000
max 8.000000 7.000000
b count 3.000000 3.000000
mean 4.000000 6.000000
std 4.358899 2.645751
min 1.000000 4.000000
25% 1.500000 4.500000
50% 2.000000 5.000000
75% 5.500000 7.000000
max 9.000000 9.000000
2. 可自定义函数，传入agg方法中
grouped.agg(func)
func的参数为groupby索引对应的记录
示例代码：
# 自定义聚合函数
def peak_range(df):
"""
返回数值范围
"""
#print type(df) #参数为索引所对应的记录
return df.max() - df.min()
print(df_obj5.groupby('key1').agg(peak_range))
print(df_obj.groupby('key1').agg(lambda df : df.max() - df.min()))
运行结果：
data1 data2
key1
a 5 3
b 8 5
data1 data2
key1
a 2.528067 1.594711
b 0.787527 0.386341
In [25]:
3. 应用多个聚合函数
同时应用多个函数进行聚合操作，使用函数列表
示例代码：
# 应用多个聚合函数
# 同时应用多个聚合函数
print(df_obj.groupby('key1').agg(['mean', 'std', 'count', peak_range])) # 默认列名为函数名
print(df_obj.groupby('key1').agg(['mean', 'std', 'count', ('range', peak_range)])) # 通过元组提供新的列名
运行结果：
data1 data2
mean std count peak_range mean std count peak_range
key1
a 0.437389 1.174151 5 2.528067 -0.230101 0.686488 5 1.594711
b 0.014657 0.440878 3 0.787527 0.802114 0.196850 3 0.386341
data1 data2
mean std count range mean std count range
key1
a 0.437389 1.174151 5 2.528067 -0.230101 0.686488 5 1.594711
b 0.014657 0.440878 3 0.787527 0.802114 0.196850 3 0.386341
4. 对不同的列分别作用不同的聚合函数，使用dict
示例代码：
# 每列作用不同的聚合函数
dict_mapping = {'data1':'mean',
'data2':'sum'}
print(df_obj.groupby('key1').agg(dict_mapping))
dict_mapping = {'data1':['mean','max'],
'data2':'sum'}
print(df_obj.groupby('key1').agg(dict_mapping))
运行结果：
data1 data2
key1
a 0.437389 -1.150505
b 0.014657 2.406341
data1 data2
mean max sum
key1
a 0.437389 1.508838 -1.150505
b 0.014657 0.522911 2.406341
5. 常用的内置聚合函数
数据的分组运算
示例代码：
import pandas as pd
import numpy as np
dict_obj = {'key1' : ['a', 'b', 'a', 'b',
'a', 'b', 'a', 'a'],
'key2' : ['one', 'one', 'two', 'three',
'two', 'two', 'one', 'three'],
'data1': np.random.randint(1, 10, 8),
'data2': np.random.randint(1, 10, 8)}
df_obj = pd.DataFrame(dict_obj)
print(df_obj)
# 按key1分组后，计算data1，data2的统计信息并附加到原始表格中，并添加表头前缀
k1_sum = df_obj.groupby('key1').sum().add_prefix('sum_')
print(k1_sum)
运行结果：
data1 data2 key1 key2
0 5 1 a one
1 7 8 b one
2 1 9 a two
3 2 6 b three
4 9 8 a two
5 8 3 b two
6 3 5 a one
7 8 3 a three
sum_data1 sum_data2
key1
a 26 26
b 17 17
聚合运算后会改变原始数据的形状，
如何保持原始数据的形状?
1. merge
使用merge的外连接，比较复杂
示例代码：
# 方法1，使用merge
k1_sum_merge = pd.merge(df_obj, k1_sum, left_on='key1', right_index=True)
print(k1_sum_merge)
运行结果：
data1 data2 key1 key2 sum_data1 sum_data2
0 5 1 a one 26 26
2 1 9 a two 26 26
4 9 8 a two 26 26
6 3 5 a one 26 26
7 8 3 a three 26 26
1 7 8 b one 17 17
3 2 6 b three 17 17
5 8 3 b two 17 17
2. transform
transform的计算结果和原始数据的形状保持一致，
如：grouped.transform(np.sum)
示例代码：
# 方法2，使用transform
k1_sum_tf = df_obj.groupby('key1').transform(np.sum).add_prefix('sum_')
df_obj[k1_sum_tf.columns] = k1_sum_tf
print(df_obj)
运行结果：
data1 data2 key1 key2 sum_data1 sum_data2 sum_key2
0 5 1 a one 26 26 onetwotwoonethree
1 7 8 b one 17 17 onethreetwo
2 1 9 a two 26 26 onetwotwoonethree
3 2 6 b three 17 17 onethreetwo
4 9 8 a two 26 26 onetwotwoonethree
5 8 3 b two 17 17 onethreetwo
6 3 5 a one 26 26 onetwotwoonethree
7 8 3 a three 26 26 onetwotwoonethree
也可传入自定义函数，
示例代码：
# 自定义函数传入transform
def diff_mean(s):
"""
返回数据与均值的差值
"""
return s - s.mean()
print(df_obj.groupby('key1').transform(diff_mean))
运行结果：
data1 data2 sum_data1 sum_data2
0 -0.200000 -4.200000 0 0
1 1.333333 2.333333 0 0
2 -4.200000 3.800000 0 0
3 -3.666667 0.333333 0 0
4 3.800000 2.800000 0 0
5 2.333333 -2.666667 0 0
6 -2.200000 -0.200000 0 0
7 2.800000 -2.200000 0 0
groupby.apply(func)
func函数也可以在各分组上分别调用，最后结果通过pd.concat组装到一起（数据合并）
示例代码：
import pandas as pd
import numpy as np
dataset_path = './starcraft.csv'
df_data = pd.read_csv(dataset_path, usecols=['LeagueIndex', 'Age', 'HoursPerWeek',
'TotalHours', 'APM'])
def top_n(df, n=3, column='APM'):
"""
返回每个分组按 column 的 top n 数据
"""
return df.sort_values(by=column, ascending=False)[:n]
print(df_data.groupby('LeagueIndex').apply(top_n))
运行结果：
LeagueIndex Age HoursPerWeek TotalHours APM
LeagueIndex
1 2214 1 20.0 12.0 730.0 172.9530
2246 1 27.0 8.0 250.0 141.6282
1753 1 20.0 28.0 100.0 139.6362
2 3062 2 20.0 6.0 100.0 179.6250
3229 2 16.0 24.0 110.0 156.7380
1520 2 29.0 6.0 250.0 151.6470
3 1557 3 22.0 6.0 200.0 226.6554
484 3 19.0 42.0 450.0 220.0692
2883 3 16.0 8.0 800.0 208.9500
4 2688 4 26.0 24.0 990.0 249.0210
1759 4 16.0 6.0 75.0 229.9122
2637 4 23.0 24.0 650.0 227.2272
5 3277 5 18.0 16.0 950.0 372.6426
93 5 17.0 36.0 720.0 335.4990
202 5 37.0 14.0 800.0 327.7218
6 734 6 16.0 28.0 730.0 389.8314
2746 6 16.0 28.0 4000.0 350.4114
1810 6 21.0 14.0 730.0 323.2506
7 3127 7 23.0 42.0 2000.0 298.7952
104 7 21.0 24.0 1000.0 286.4538
1654 7 18.0 98.0 700.0 236.0316
8 3393 8 NaN NaN NaN 375.8664
3373 8 NaN NaN NaN 364.8504
3372 8 NaN NaN NaN 355.3518
1. 产生层级索引：外层索引是分组名，内层索引是df_obj的行索引
示例代码：
# apply函数接收的参数会传入自定义的函数中
print(df_data.groupby('LeagueIndex').apply(top_n, n=2, column='Age'))
运行结果：
LeagueIndex Age HoursPerWeek TotalHours APM
LeagueIndex
1 3146 1 40.0 12.0 150.0 38.5590
3040 1 39.0 10.0 500.0 29.8764
2 920 2 43.0 10.0 730.0 86.0586
2437 2 41.0 4.0 200.0 54.2166
3 1258 3 41.0 14.0 800.0 77.6472
2972 3 40.0 10.0 500.0 60.5970
4 1696 4 44.0 6.0 500.0 89.5266
1729 4 39.0 8.0 500.0 86.7246
5 202 5 37.0 14.0 800.0 327.7218
2745 5 37.0 18.0 1000.0 123.4098
6 3069 6 31.0 8.0 800.0 133.1790
2706 6 31.0 8.0 700.0 66.9918
7 2813 7 26.0 36.0 1300.0 188.5512
1992 7 26.0 24.0 1000.0 219.6690
8 3340 8 NaN NaN NaN 189.7404
3341 8 NaN NaN NaN 287.8128
2. 禁止层级索引, group_keys=False
示例代码：
print(df_data.groupby('LeagueIndex', group_keys=False).apply(top_n))
运行结果：
LeagueIndex Age HoursPerWeek TotalHours APM
2214 1 20.0 12.0 730.0 172.9530
2246 1 27.0 8.0 250.0 141.6282
1753 1 20.0 28.0 100.0 139.6362
3062 2 20.0 6.0 100.0 179.6250
3229 2 16.0 24.0 110.0 156.7380
1520 2 29.0 6.0 250.0 151.6470
1557 3 22.0 6.0 200.0 226.6554
484 3 19.0 42.0 450.0 220.0692
2883 3 16.0 8.0 800.0 208.9500
2688 4 26.0 24.0 990.0 249.0210
1759 4 16.0 6.0 75.0 229.9122
2637 4 23.0 24.0 650.0 227.2272
3277 5 18.0 16.0 950.0 372.6426
93 5 17.0 36.0 720.0 335.4990
202 5 37.0 14.0 800.0 327.7218
734 6 16.0 28.0 730.0 389.8314
2746 6 16.0 28.0 4000.0 350.4114
1810 6 21.0 14.0 730.0 323.2506
3127 7 23.0 42.0 2000.0 298.7952
104 7 21.0 24.0 1000.0 286.4538
1654 7 18.0 98.0 700.0 236.0316
3393 8 NaN NaN NaN 375.8664
3373 8 NaN NaN NaN 364.8504
3372 8 NaN NaN NaN 355.3518
apply可以用来处理不同分组内的缺失数据填充，填充该分组的均值。