目录
我们在这里要讲一个很常用的技术, 就是所谓的分组技术, 这个在数据库中是非常常用的, 要去求某些分组的统计量, 那么我们需要知道在pandas里面, 这些分组技术是怎么实现的.
分组操作
我们这里要来聊聊在pandas中实现分组运算, 大致上可以按照列, 字典或者Series, 函数, 索引级别进行分组, 我们会逐渐来介绍.
按照列进行分组
import pandas as pd
from pandas import DataFrame, Series
import numpy as np
sep = "---------------------------------------------------------------------------"
data = DataFrame({"key1": [‘a‘, ‘a‘, ‘b‘, ‘b‘, ‘a‘], "key2": [‘one‘, ‘two‘, ‘one‘, ‘two‘, ‘one‘], ‘data1‘: np.random.randn(5), ‘data2‘: np.random.randn(5)})
print(data)
data1 data2 key1 key2
0 0.733951 0.000379 a one
1 1.039029 0.852930 a two
2 0.921413 -1.644942 b one
3 0.294560 0.521525 b two
4 0.286072 -0.074574 a one
data1按照key1分组为:
groups = data[‘data1‘].groupby(data[‘key1‘])
我们发现得到了一个SeriesGroupBy 对象, 现在我们对这个对象进行迭代:
for name, group in groups:
print(name)
print(sep)
print(group)
print(sep)
a
---------------------------------------------------------------------------
0 0.733951
1 1.039029
4 0.286072
Name: data1, dtype: float64
---------------------------------------------------------------------------
b
---------------------------------------------------------------------------
2 0.921413
3 0.294560
Name: data1, dtype: float64
---------------------------------------------------------------------------
我们发现, groups有(key, Series)对组成, key根据什么来分组的元素, Series(DataFrame)是分组的元素, Series(DataFrame)的name还是原来的列名.
对你分组进行迭代, 用:
for name, group in groups
groups = data.groupby(data[‘key1‘])
for name, group in groups:
print(name)
print(sep)
print(group)
print(sep)
a
---------------------------------------------------------------------------
data1 data2 key1 key2
0 0.733951 0.000379 a one
1 1.039029 0.852930 a two
4 0.286072 -0.074574 a one
---------------------------------------------------------------------------
b
---------------------------------------------------------------------------
data1 data2 key1 key2
2 0.921413 -1.644942 b one
3 0.294560 0.521525 b two
---------------------------------------------------------------------------
groupby就是按照某个值来分组, 无论是对series还是dataframe, 都成立.
我们可以在分好组的对象上调用统计函数.
data.groupby(data[‘key1‘]).mean()
data1 | data2 | |
---|---|---|
key1 | ||
a | 0.686351 | 0.259578 |
b | 0.607986 | -0.561709 |
在每个分组上分别对每个每一列求均值, 如果是非数字列, 或默认剔除.
作业1:在每个分组上分别对每个每一行求均值.
提示: data.groupby(data[‘key1‘]).mean(axis=1)是行不通的.
对于多个列进行分组, 分组的key是对应分组元素的元组.
作业2:对DataFrame用多个列进行分组.
下面其我们来看一个语法糖:
data.groupby([data[‘key1‘], data[‘key2‘]])
<pandas.core.groupby.DataFrameGroupBy object at 0x000001D080230278>
它等价于:
data.groupby([‘key1‘, ‘key2‘])
<pandas.core.groupby.DataFrameGroupBy object at 0x000001D080230630>
我们来验证一下:
groups =data.groupby([data[‘key1‘], data[‘key2‘]])
for name, group in groups:
print(name)
print(sep)
print(group)
print(sep)
(‘a‘, ‘one‘)
---------------------------------------------------------------------------
data1 data2 key1 key2
0 0.733951 0.000379 a one
4 0.286072 -0.074574 a one
---------------------------------------------------------------------------
(‘a‘, ‘two‘)
---------------------------------------------------------------------------
data1 data2 key1 key2
1 1.039029 0.85293 a two
---------------------------------------------------------------------------
(‘b‘, ‘one‘)
---------------------------------------------------------------------------
data1 data2 key1 key2
2 0.921413 -1.644942 b one
---------------------------------------------------------------------------
(‘b‘, ‘two‘)
---------------------------------------------------------------------------
data1 data2 key1 key2
3 0.29456 0.521525 b two
---------------------------------------------------------------------------
groups = data.groupby([‘key1‘, ‘key2‘])
for name, group in groups:
print(name)
print(sep)
print(group)
print(sep)
(‘a‘, ‘one‘)
---------------------------------------------------------------------------
data1 data2 key1 key2
0 0.733951 0.000379 a one
4 0.286072 -0.074574 a one
---------------------------------------------------------------------------
(‘a‘, ‘two‘)
---------------------------------------------------------------------------
data1 data2 key1 key2
1 1.039029 0.85293 a two
---------------------------------------------------------------------------
(‘b‘, ‘one‘)
---------------------------------------------------------------------------
data1 data2 key1 key2
2 0.921413 -1.644942 b one
---------------------------------------------------------------------------
(‘b‘, ‘two‘)
---------------------------------------------------------------------------
data1 data2 key1 key2
3 0.29456 0.521525 b two
---------------------------------------------------------------------------
我们发现输出结果是一模一样, 总结一下:
data.groupby([data[‘key1‘], data[‘key2‘]])等价于data.groupby([‘key1‘, ‘key2‘])
进一步:
data[‘data1‘].groupby([data[‘key1‘], data[‘key2‘]])等价于data.groupby([‘key1‘, ‘key2‘])[‘data1‘]
作业3: 验证data[‘data1‘].groupby([data[‘key1‘], data[‘key2‘]])等价于data.groupby([‘key1‘, ‘key2‘])[‘data1‘].
data.groupby([‘key1‘, ‘key2‘])[‘data1‘]
<pandas.core.groupby.SeriesGroupBy object at 0x000001D0FCD95D68>
data.groupby([‘key1‘, ‘key2‘])[[‘data1‘]]
<pandas.core.groupby.DataFrameGroupBy object at 0x000001D080232898>
我不知道大家发现没有, 这两个返回的数据类型是有区别的, 我们仔细来看看:
data[[‘data1‘]] # 这是一个DataFrame
data1 | |
---|---|
0 | 0.733951 |
1 | 1.039029 |
2 | 0.921413 |
3 | 0.294560 |
4 | 0.286072 |
data[‘data1‘] # 这是一个Series
0 0.733951
1 1.039029
2 0.921413
3 0.294560
4 0.286072
Name: data1, dtype: float64
那么这里的区别就不言而喻了吧
groups = data.groupby([‘key1‘, ‘key2‘])[[‘data1‘]]
for name, group in groups:
print(name)
print(sep)
print(group)
print(sep)
(‘a‘, ‘one‘)
---------------------------------------------------------------------------
<class ‘pandas.core.frame.DataFrame‘>
---------------------------------------------------------------------------
(‘a‘, ‘two‘)
---------------------------------------------------------------------------
<class ‘pandas.core.frame.DataFrame‘>
---------------------------------------------------------------------------
(‘b‘, ‘one‘)
---------------------------------------------------------------------------
<class ‘pandas.core.frame.DataFrame‘>
---------------------------------------------------------------------------
(‘b‘, ‘two‘)
---------------------------------------------------------------------------
<class ‘pandas.core.frame.DataFrame‘>
---------------------------------------------------------------------------
结果是一样的.
data.groupby([‘key1‘, ‘key2‘])[[‘data1‘]].mean()
data1 | ||
---|---|---|
key1 | key2 | |
a | one | 0.510012 |
two | 1.039029 | |
b | one | 0.921413 |
two | 0.294560 |
data.groupby([‘key1‘, ‘key2‘])[‘data1‘].mean()
key1 key2
a one 0.510012
two 1.039029
b one 0.921413
two 0.294560
Name: data1, dtype: float64
在做数据聚合的时候就发现了不同,
[[‘data1‘]]得到的是一个DataFrame, 而[‘data1‘]得到的是Series.
按照字典进行分组
我们来看一个按照字典进行分组的例子:
data = DataFrame(np.random.randn(5, 5), columns=[‘a‘, ‘b‘, ‘c‘, ‘d‘, ‘e‘], index=[‘joe‘, ‘steve‘, ‘wes‘, ‘jim‘, ‘Travis‘])
data
a | b | c | d | e | |
---|---|---|---|---|---|
joe | -0.089597 | 1.239307 | 2.173063 | -0.519295 | -1.783812 |
steve | 0.539109 | 0.724553 | -0.041899 | 0.787494 | 0.394633 |
wes | -0.055417 | 0.384068 | -0.594006 | -0.451587 | 0.722761 |
jim | -0.056767 | 0.398863 | 2.140669 | -1.060791 | -0.953756 |
Travis | 0.245142 | -0.468819 | -0.863372 | -0.151966 | 1.185567 |
# 定义一个分组的字典, a, b, c --> red, d, e --> blue
mapping = {‘a‘:‘red‘, ‘b‘:‘red‘, ‘c‘: ‘red‘, ‘d‘:‘blue‘, ‘e‘: ‘blue‘}
data.groupby(mapping, axis=1).mean() # 对每一个分组求平均
blue | red | |
---|---|---|
joe | -1.151554 | 1.107591 |
steve | 0.591063 | 0.407255 |
wes | 0.135587 | -0.088452 |
jim | -1.007273 | 0.827589 |
Travis | 0.516800 | -0.362350 |
作业4:自己设计一个index的mapping, 按axis=0进行分组.
根据函数进行分组
话不多说, 直接来看例子:
data.groupby(len).mean()
a | b | c | d | e | |
---|---|---|---|---|---|
3 | -0.067260 | 0.674079 | 1.239909 | -0.677224 | -0.671602 |
5 | 0.539109 | 0.724553 | -0.041899 | 0.787494 | 0.394633 |
6 | 0.245142 | -0.468819 | -0.863372 | -0.151966 | 1.185567 |
我们发现, 字典和函数都是作用到索引上的.
按照list组合
这个例子非常简单:
data.groupby([‘1‘, ‘1‘, ‘1‘, ‘2‘, ‘2‘]).mean()
a | b | c | d | e | |
---|---|---|---|---|---|
1 | 0.131365 | 0.782643 | 0.512386 | -0.061130 | -0.222139 |
2 | 0.094188 | -0.034978 | 0.638649 | -0.606378 | 0.115905 |
他会自动判断是按照列还是list.
按照索引级别进行分组
作业5: 自己学习按索引级别进行分组.
分组运算
分组运算主要设计到3个函数, agg, transform和apply.
我们一个一个来看.
agg
data = DataFrame({"key1": [‘a‘, ‘a‘, ‘b‘, ‘b‘, ‘a‘], "key2": [‘one‘, ‘two‘, ‘one‘, ‘two‘, ‘one‘], ‘data1‘: np.random.randn(5), ‘data2‘: np.random.randn(5)})
data
data1 | data2 | key1 | key2 | |
---|---|---|---|---|
0 | 0.441278 | -0.848457 | a | one |
1 | 1.843375 | -0.522482 | a | two |
2 | -1.435176 | -0.191682 | b | one |
3 | -2.700772 | -0.832993 | b | two |
4 | -1.430386 | -1.910834 | a | one |
data.groupby("key2").agg(np.mean)
data1 | data2 | |
---|---|---|
key2 | ||
one | -0.808095 | -0.983658 |
two | -0.428699 | -0.677738 |
当然, 这个等价于:
data.groupby("key2").mean()
data1 | data2 | |
---|---|---|
key2 | ||
one | -0.808095 | -0.983658 |
two | -0.428699 | -0.677738 |
原理: 聚合函数会在group后的每个切片上(group后的每一行或每一列)计算出值.
我们也可以自定义函数:
data.groupby("key2").agg(lambda x: x.max() - x.min())
data1 | data2 | |
---|---|---|
key2 | ||
one | 1.876454 | 1.719153 |
two | 4.544147 | 0.310511 |
他会在每个分组的每个列上调用这个函数.
data.groupby("key2").agg([np.mean, np.max,np.min])
data1 | data2 | |||||
---|---|---|---|---|---|---|
mean | amax | amin | mean | amax | amin | |
key2 | ||||||
one | -0.808095 | 0.441278 | -1.435176 | -0.983658 | -0.191682 | -1.910834 |
two | -0.428699 | 1.843375 | -2.700772 | -0.677738 | -0.522482 | -0.832993 |
data.groupby("key2").agg([("平均值:", np.mean), ("最大值",np.max), ("最小值",np.min)]).rename({"one": "第一组", "two":"第二组"})
data1 | data2 | |||||
---|---|---|---|---|---|---|
平均值: | 最大值 | 最小值 | 平均值: | 最大值 | 最小值 | |
key2 | ||||||
第一组 | -0.808095 | 0.441278 | -1.435176 | -0.983658 | -0.191682 | -1.910834 |
第二组 | -0.428699 | 1.843375 | -2.700772 | -0.677738 | -0.522482 | -0.832993 |
# 对不同列用不同的分组函数
data.groupby("key2").agg({"data1":[("平均值:", np.mean), ("最大值",np.max)], "data2":[("最小值",np.min)]}).rename({"one": "第一组", "two":"第二组"})
data2 | data1 | ||
---|---|---|---|
最小值 | 平均值: | 最大值 | |
key2 | |||
第一组 | -1.910834 | -0.808095 | 0.441278 |
第二组 | -0.832993 | -0.428699 | 1.843375 |
transform
transform是一个矢量化的函数, 如果最后我们得到的值和分组切片不一致, 会进行广播:
data
data1 | data2 | key1 | key2 | |
---|---|---|---|---|
0 | 0.441278 | -0.848457 | a | one |
1 | 1.843375 | -0.522482 | a | two |
2 | -1.435176 | -0.191682 | b | one |
3 | -2.700772 | -0.832993 | b | two |
4 | -1.430386 | -1.910834 | a | one |
data.groupby("key1").transform(np.mean)
data1 | data2 | |
---|---|---|
0 | 0.284756 | -1.093924 |
1 | 0.284756 | -1.093924 |
2 | -2.067974 | -0.512338 |
3 | -2.067974 | -0.512338 |
4 | 0.284756 | -1.093924 |
仔细看, 0,1, 4一组, 2,3一组, 发生了广播.
现在有个需求,按分组减去均值.
data.groupby("key1").transform(lambda x: x - x.mean())
data1 | data2 | |
---|---|---|
0 | 0.156523 | 0.245468 |
1 | 1.558619 | 0.571442 |
2 | 0.632798 | 0.320656 |
3 | -0.632798 | -0.320656 |
4 | -1.715142 | -0.816910 |
a, b分组的各列都减去了他们的均值, 不信, 来看:
data.groupby("key1").transform(lambda x: x - x.mean()).groupby([1, 1, 0,0, 1]).mean()
data1 | data2 | |
---|---|---|
0 | 1.110223e-16 | -5.551115e-17 |
1 | 7.401487e-17 | -1.110223e-16 |
apply
这个函数是transform的加强版, transform只能返回和原来切片大小一样大的, 但apply是可以任意的. 其实我们之前就用过apply函数, 我们知道, apply是作用在列(行)上的, applymap是作用在函数上的.
data = DataFrame({"key1": [‘a‘, ‘a‘, ‘b‘, ‘b‘, ‘a‘], "key2": [‘one‘, ‘two‘, ‘one‘, ‘two‘, ‘one‘], ‘data1‘: np.random.randn(5), ‘data2‘: np.random.randn(5)})
data
data1 | data2 | key1 | key2 | |
---|---|---|---|---|
0 | -0.312694 | 0.073574 | a | one |
1 | -0.902065 | -0.854249 | a | two |
2 | -0.440915 | 0.228551 | b | one |
3 | -0.406243 | -0.878505 | b | two |
4 | 1.812926 | -0.114598 | a | one |
如果我们要找出one, 和two分组中选出data2最大的前两个呢?
data.groupby(‘key2‘).apply(lambda x: x.sort_values(by=‘data2‘)[-2:])
data1 | data2 | key1 | key2 | ||
---|---|---|---|---|---|
key2 | |||||
one | 0 | -0.312694 | 0.073574 | a | one |
2 | -0.440915 | 0.228551 | b | one | |
two | 3 | -0.406243 | -0.878505 | b | two |
1 | -0.902065 | -0.854249 | a | two |
去掉group层次索引:
data.groupby(‘key2‘, group_keys=False).apply(lambda x: x.sort_values(by=‘data2‘)[-2:])
data1 | data2 | key1 | key2 | |
---|---|---|---|---|
0 | -0.312694 | 0.073574 | a | one |
2 | -0.440915 | 0.228551 | b | one |
3 | -0.406243 | -0.878505 | b | two |
1 | -0.902065 | -0.854249 | a | two |
总结一下: apply就是把分完组的切片挨个(按行, 按列, 或者整体)调用我们的函数, 最后再把结果合并起来.
利用groupby技术多进程处理DataFrame
我们这里要教大家用一种groupby技术, 来实现对DataFrame并行处理.
pip install joblib
因为我们windows系统的限制, 我们的代码是在linux上运行的:
import math
from joblib import Parallel, delayed
from pandas import DataFrame
import pandas as pd
import numpy as np
import time
begin = time.time()
test = DataFrame(np.random.randn(10000000, 10))
test_other = test.copy()
groups = test.groupby(lambda x: x % 8)
def func(x):
return x.applymap(lambda y: math.pow(y, 4))
pd.concat(Parallel(n_jobs=8)(delayed(func)(group) for name, group in groups))
print(time.time() - begin)
begin = time.time()
test_other.applymap(lambda x: math.pow(x, 4))
print(time.time() - begin)
运算结果为:
23.35878014564514
62.76386260986328
速度大概提升了2.5倍, 还是很不错的.