单个变量的频率表

问题:单个变量的频率表

当天最后一个新手熊猫问题:如何为单个系列生成一张表?

例如:

my_series = pandas.Series([1,2,2,3,3,3])
pandas.magical_frequency_function( my_series )

>> {
     1 : 1,
     2 : 2, 
     3 : 3
   }

大量的搜索使我进入了Series.describe()和pandas.crosstabs,但是这些都不满足我的需要:一个变量,按类别计数。哦,如果它适用于不同的数据类型(字符串,整数等),那就太好了。

One last newbie pandas question for the day: How do I generate a table for a single Series?

For example:

my_series = pandas.Series([1,2,2,3,3,3])
pandas.magical_frequency_function( my_series )

>> {
     1 : 1,
     2 : 2, 
     3 : 3
   }

Lots of googling has led me to Series.describe() and pandas.crosstabs, but neither of these does quite what I need: one variable, counts by categories. Oh, and it’d be nice if it worked for different data types: strings, ints, etc.


回答 0

也许.value_counts()吧?

>>> import pandas
>>> my_series = pandas.Series([1,2,2,3,3,3, "fred", 1.8, 1.8])
>>> my_series
0       1
1       2
2       2
3       3
4       3
5       3
6    fred
7     1.8
8     1.8
>>> counts = my_series.value_counts()
>>> counts
3       3
2       2
1.8     2
fred    1
1       1
>>> len(counts)
5
>>> sum(counts)
9
>>> counts["fred"]
1
>>> dict(counts)
{1.8: 2, 2: 2, 3: 3, 1: 1, 'fred': 1}

Maybe .value_counts()?

>>> import pandas
>>> my_series = pandas.Series([1,2,2,3,3,3, "fred", 1.8, 1.8])
>>> my_series
0       1
1       2
2       2
3       3
4       3
5       3
6    fred
7     1.8
8     1.8
>>> counts = my_series.value_counts()
>>> counts
3       3
2       2
1.8     2
fred    1
1       1
>>> len(counts)
5
>>> sum(counts)
9
>>> counts["fred"]
1
>>> dict(counts)
{1.8: 2, 2: 2, 3: 3, 1: 1, 'fred': 1}

回答 1

您可以对数据框使用列表理解来计算列的频率,例如

[my_series[c].value_counts() for c in list(my_series.select_dtypes(include=['O']).columns)]

分解:

my_series.select_dtypes(include=['O']) 

仅选择分类数据

list(my_series.select_dtypes(include=['O']).columns) 

从上方将列转换为列表

[my_series[c].value_counts() for c in list(my_series.select_dtypes(include=['O']).columns)] 

遍历上面的列表,并将value_counts()应用于每个列

You can use list comprehension on a dataframe to count frequencies of the columns as such

[my_series[c].value_counts() for c in list(my_series.select_dtypes(include=['O']).columns)]

Breakdown:

my_series.select_dtypes(include=['O']) 

Selects just the categorical data

list(my_series.select_dtypes(include=['O']).columns) 

Turns the columns from above into a list

[my_series[c].value_counts() for c in list(my_series.select_dtypes(include=['O']).columns)] 

Iterates through the list above and applies value_counts() to each of the columns


回答 2

@DSM提供的答案很简单明了,但是我想我将自己的输入添加到该问题中。如果查看pandas.value_counts的代码,将会发现发生了很多事情。

如果您需要计算多个序列的频率,则可能需要一段时间。更快的实现是将numpy.uniquereturn_counts = True

这是一个例子:

import pandas as pd
import numpy as np

my_series = pd.Series([1,2,2,3,3,3])

print(my_series.value_counts())
3    3
2    2
1    1
dtype: int64

注意这里返回的项目是pandas.Series

相比之下,numpy.unique返回具有两个项目的元组,即唯一值和计数。

vals, counts = np.unique(my_series, return_counts=True)
print(vals, counts)
[1 2 3] [1 2 3]

然后,您可以将它们组合成字典:

results = dict(zip(vals, counts))
print(results)
{1: 1, 2: 2, 3: 3}

然后变成 pandas.Series

print(pd.Series(results))
1    1
2    2
3    3
dtype: int64

The answer provided by @DSM is simple and straightforward, but I thought I’d add my own input to this question. If you look at the code for pandas.value_counts, you’ll see that there is a lot going on.

If you need to calculate the frequency of many series, this could take a while. A faster implementation would be to use numpy.unique with return_counts = True

Here is an example:

import pandas as pd
import numpy as np

my_series = pd.Series([1,2,2,3,3,3])

print(my_series.value_counts())
3    3
2    2
1    1
dtype: int64

Notice here that the item returned is a pandas.Series

In comparison, numpy.unique returns a tuple with two items, the unique values and the counts.

vals, counts = np.unique(my_series, return_counts=True)
print(vals, counts)
[1 2 3] [1 2 3]

You can then combine these into a dictionary:

results = dict(zip(vals, counts))
print(results)
{1: 1, 2: 2, 3: 3}

And then into a pandas.Series

print(pd.Series(results))
1    1
2    2
3    3
dtype: int64

回答 3

对于具有过多值的变量的频率分布,您可以将类中的值折叠起来,

在这里,我为employrate变量设置了过多的值,并且没有直接频率分布的含义values_count(normalize=True)

                country  employrate alcconsumption
0           Afghanistan   55.700001            .03
1               Albania   11.000000           7.29
2               Algeria   11.000000            .69
3               Andorra         nan          10.17
4                Angola   75.699997           5.57
..                  ...         ...            ...
208             Vietnam   71.000000           3.91
209  West Bank and Gaza   32.000000               
210         Yemen, Rep.   39.000000             .2
211              Zambia   61.000000           3.56
212            Zimbabwe   66.800003           4.96

[213 rows x 3 columns]

values_count(normalize=True)没有分类的频率分布,结果长度为139(似乎无意义的频率分布):

print(gm["employrate"].value_counts(sort=False,normalize=True))

50.500000   0.005618
61.500000   0.016854
46.000000   0.011236
64.500000   0.005618
63.500000   0.005618

58.599998   0.005618
63.799999   0.011236
63.200001   0.005618
65.599998   0.005618
68.300003   0.005618
Name: employrate, Length: 139, dtype: float64

进行分类时,我们将所有值都放在一定范围内。

0-10为1
11-20为2  
21-30为3,依此类推。
gm["employrate"]=gm["employrate"].str.strip().dropna()  
gm["employrate"]=pd.to_numeric(gm["employrate"])
gm['employrate'] = np.where(
   (gm['employrate'] <=10) & (gm['employrate'] > 0) , 1, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=20) & (gm['employrate'] > 10) , 1, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=30) & (gm['employrate'] > 20) , 2, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=40) & (gm['employrate'] > 30) , 3, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=50) & (gm['employrate'] > 40) , 4, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=60) & (gm['employrate'] > 50) , 5, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=70) & (gm['employrate'] > 60) , 6, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=80) & (gm['employrate'] > 70) , 7, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=90) & (gm['employrate'] > 80) , 8, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=100) & (gm['employrate'] > 90) , 9, gm['employrate']
   )
print(gm["employrate"].value_counts(sort=False,normalize=True))

分类后,我们有一个清晰的频率分布。在这里我们可以很容易地看到,37.64%一个国家的雇佣率介于51-60%11.79%之间。71-80%

5.000000   0.376404
7.000000   0.117978
4.000000   0.179775
6.000000   0.264045
8.000000   0.033708
3.000000   0.028090
Name: employrate, dtype: float64

for frequency distribution of a variable with excessive values you can collapse down the values in classes,

Here I excessive values for employrate variable, and there’s no meaning of it’s frequency distribution with direct values_count(normalize=True)

                country  employrate alcconsumption
0           Afghanistan   55.700001            .03
1               Albania   11.000000           7.29
2               Algeria   11.000000            .69
3               Andorra         nan          10.17
4                Angola   75.699997           5.57
..                  ...         ...            ...
208             Vietnam   71.000000           3.91
209  West Bank and Gaza   32.000000               
210         Yemen, Rep.   39.000000             .2
211              Zambia   61.000000           3.56
212            Zimbabwe   66.800003           4.96

[213 rows x 3 columns]

frequency distribution with values_count(normalize=True) with no classification,length of result here is 139 (seems meaningless as a frequency distribution):

print(gm["employrate"].value_counts(sort=False,normalize=True))

50.500000   0.005618
61.500000   0.016854
46.000000   0.011236
64.500000   0.005618
63.500000   0.005618

58.599998   0.005618
63.799999   0.011236
63.200001   0.005618
65.599998   0.005618
68.300003   0.005618
Name: employrate, Length: 139, dtype: float64

putting classification we put all values with a certain range ie.

0-10 as 1,
11-20 as 2  
21-30 as 3, and so forth.
gm["employrate"]=gm["employrate"].str.strip().dropna()  
gm["employrate"]=pd.to_numeric(gm["employrate"])
gm['employrate'] = np.where(
   (gm['employrate'] <=10) & (gm['employrate'] > 0) , 1, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=20) & (gm['employrate'] > 10) , 1, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=30) & (gm['employrate'] > 20) , 2, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=40) & (gm['employrate'] > 30) , 3, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=50) & (gm['employrate'] > 40) , 4, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=60) & (gm['employrate'] > 50) , 5, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=70) & (gm['employrate'] > 60) , 6, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=80) & (gm['employrate'] > 70) , 7, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=90) & (gm['employrate'] > 80) , 8, gm['employrate']
   )
gm['employrate'] = np.where(
   (gm['employrate'] <=100) & (gm['employrate'] > 90) , 9, gm['employrate']
   )
print(gm["employrate"].value_counts(sort=False,normalize=True))

after classification we have a clear frequency distribution. here we can easily see, that 37.64% of countries have employ rate between 51-60% and 11.79% of countries have employ rate between 71-80%

5.000000   0.376404
7.000000   0.117978
4.000000   0.179775
6.000000   0.264045
8.000000   0.033708
3.000000   0.028090
Name: employrate, dtype: float64