本文目标：

1. 应用groupby 进行分组
2. 对分组数据进行聚合,转换和过滤
3. 应用自定义函数处理分组之后的数据

---

1. 数据聚合

• 在SQL中我们经常使用 GROUP BY 将某个字段,按不同的取值进行分组, 在pandas中也有groupby函数
• 分组之后,每组都会有至少1条数据, 将这些数据进一步处理返回单个值的过程就是聚合,比如分组之后计算算术平均值, 或者分组之后计算频数,都属于聚合

1.1 单变量分组聚合

# 数据片段
'''
       country continent  year  lifeExp       pop   gdpPercap
0  Afghanistan      Asia  1952   28.801   8425333  779.445314
1  Afghanistan      Asia  1957   30.332   9240934  820.853030
2  Afghanistan      Asia  1962   31.997  10267083  853.100710
3  Afghanistan      Asia  1967   34.020  11537966  836.197138
4  Afghanistan      Asia  1972   36.088  13079460  739.981106
'''
df = pd.read_csv('data/gapminder.tsv', sep='\t')
# 单变量分组聚合
df.groupby('year').lifeExp.mean()

'''
year
1952    49.057620
1957    51.507401
1962    53.609249
1967    55.678290
1972    57.647386
1977    59.570157
Name: lifeExp, dtype: float64
'''

1.2 Pandas内置聚合方法

1. 可以与groupby一起使用的方法和函数
   

2. 分组之后取平均也可以使用describe函数同时计算多个统计量

df.groupby('year').lifeExp.describe()
'''
      count       mean        std     min       25%      50%       75%     max
year                                                                          
1952  142.0  49.057620  12.225956  28.801  39.05900  45.1355  59.76500  72.670
1957  142.0  51.507401  12.231286  30.332  41.24750  48.3605  63.03675  73.470
1962  142.0  53.609249  12.097245  31.997  43.46850  50.8810  65.23450  73.680
1967  142.0  55.678290  11.718858  34.020  46.03375  53.8250  67.41950  74.160
1972  142.0  57.647386  11.381953  35.400  48.50025  56.5300  69.24750  74.720
1977  142.0  59.570157  11.227229  31.220  50.47550  59.6720  70.38250  76.110
1982  142.0  61.533197  10.770618  38.445  52.94000  62.4415  70.92125  77.110
1987  142.0  63.212613  10.556285  39.906  54.94075  65.8340  71.87725  78.670
1992  142.0  64.160338  11.227380  23.599  56.12175  67.7030  72.58250  79.360
1997  142.0  65.014676  11.559439  36.087  55.63375  69.3940  74.16975  80.690
2002  142.0  65.694923  12.279823  39.193  55.52225  70.8255  75.45925  82.000
2007  142.0  67.007423  12.073021  39.613  57.16025  71.9355  76.41325  82.603
'''

1.3 聚合方法

使用Numpy的聚合方法

df.groupby('year').lifeExp.agg(np.mean)
# df.groupby('year').lifeExp.aggregate(np.mean)
'''
year
1952    49.057620
1957    51.507401
1962    53.609249
1967    55.678290
1972    57.647386
Name: lifeExp, dtype: float64
'''

agg()和aggregate()是一样的

自定义方法

计算每年的平均年纪：

def my_mean(values):
    n = len(values)
    sum = 0
    for v in values:
        sum += v
    return (sum / n)
 df.groupby('year').lifeExp.agg(my_mean)

多个参数的自定义方法：

def my_mean_diff(values,diff):
    n = len(values)
    sum = 0
    for v in values:
        sum+=v
    mean = sum/n
    return mean - diff

df.groupby('year').lifeExp.agg(my_mean_diff,diff=1)

同时计算多种特征

import numpy as np
df.groupby('year').lifeExp.agg([np.count_nonzero, np.mean, np.std])
'''
      count_nonzero       mean        std
year                                     
1952            142  49.057620  12.225956
1957            142  51.507401  12.231286
1962            142  53.609249  12.097245
1967            142  55.678290  11.718858
1972            142  57.647386  11.381953
1977            142  59.570157  11.227229
1982            142  61.533197  10.770618
1987            142  63.212613  10.556285
1992            142  64.160338  11.227380
1997            142  65.014676  11.559439
2002            142  65.694923  12.279823
2007            142  67.007423  12.073021
'''

向agg/aggregate传入字典

分别对分组后的不同列使用不同聚合方法：

df.groupby('year').agg(
    {
        'lifeExp': 'mean',
        'pop': 'median',
        'gdpPercap': 'median'
    }
)
'''
        lifeExp         pop    gdpPercap
year                                    
1952  49.057620   3943953.0  1968.528344
1957  51.507401   4282942.0  2173.220291
1962  53.609249   4686039.5  2335.439533
1967  55.678290   5170175.5  2678.334740
1972  57.647386   5877996.5  3339.129407
1977  59.570157   6404036.5  3798.609244
1982  61.533197   7007320.0  4216.228428
1987  63.212613   7774861.5  4280.300366
1992  64.160338   8688686.5  4386.085502
1997  65.014676   9735063.5  4781.825478
2002  65.694923  10372918.5  5319.804524
2007  67.007423  10517531.0  6124.371108
'''

一步到位，把计算后的数据列进行命名：

df.groupby('year').agg({
    'lifeExp':'mean',
    'pop':'median',
    'gdpPercap':'median'
}).rename(columns={
    'lifeExp':'平均寿命',
    'pop':'人口中位数',
    'gdpPercap':'人均GDP中位数'
})

'''
       平均寿命       人口中位数     人均GDP中位数
year                                    
1952  49.057620   3943953.0  1968.528344
1957  51.507401   4282942.0  2173.220291
1962  53.609249   4686039.5  2335.439533
1967  55.678290   5170175.5  2678.334740
1972  57.647386   5877996.5  3339.129407
1977  59.570157   6404036.5  3798.609244
1982  61.533197   7007320.0  4216.228428
1987  63.212613   7774861.5  4280.300366
1992  64.160338   8688686.5  4386.085502
1997  65.014676   9735063.5  4781.825478
2002  65.694923  10372918.5  5319.804524
2007  67.007423  10517531.0  6124.371108
'''

2. 数据转换

• transform 需要把DataFrame中的值传递给一个函数， 而后由该函数"转换"数据。
• aggregate(聚合) 返回单个聚合值，但transform 不会减少数据量

def zscore(x):
    return (x-x.mean())/x.std()
df.groupby('year').lifeExp.transform(zscore)
'''
0      -1.656854
1      -1.731249
2      -1.786543
3      -1.848157
4      -1.894173
          ...   
1699   -0.081621
1700   -0.336974
1701   -1.574962
1702   -2.093346
1703   -1.948180
Name: lifeExp, Length: 1704, dtype: float64
'''

使用Transform之后，产生的结果和原数据的数量是一样的。

使用Transform，可以对缺失值进行填充：

def fill_na_mean(bills):
   return bills.fillna(bills.mean())
    
tips_10.groupby('sex')['total_bill'].transform(fill_na_mean)

3. 数据过滤

tips['size'].value_counts()
'''
2    156
3     38
4     37
5      5
1      4
6      4
Name: size, dtype: int64
'''
tips_filtered =tips.groupby('size').filter(lambda x:x['size'].count()>30)
tips_filtered['size'].value_counts()
'''
2    156
3     38
4     37
Name: size, dtype: int64
'''