In [10]: df = pd.DataFrame({'key1' : ['a', 'a', 'b', 'b', 'a'],
....: 'key2' : ['one', 'two', 'one', 'two', 'one'],
....: 'data1' : np.random.randn(5),
....: 'data2' : np.random.randn(5)})
In [11]: df
Out[11]:
data1 data2 key1 key2
0 -0.204708 1.393406 a one
1 0.478943 0.092908 a two
2 -0.519439 0.281746 b one
3 -0.555730 0.769023 b two
4 1.965781 1.246435 a one
In [15]: means = df['data1'].groupby([df['key1'], df['key2']]).mean()
In [16]: means
Out[16]:
key1 key2
a one 0.880536
two 0.478943
b one -0.519439
two -0.555730
Name: data1, dtype: float64
这里,我通过两个键对数据进行了分组,得到的Series具有一个层次化索引(由唯一的键对组成):
In [17]: means.unstack()
Out[17]:
key2 one two
key1
a 0.880536 0.478943
b -0.519439 -0.555730
在这个例子中,分组键均为Series。实际上,分组键可以是任何长度适当的数组:
In [18]: states = np.array(['Ohio', 'California', 'California', 'Ohio', 'Ohio'])
In [19]: years = np.array([2005, 2005, 2006, 2005, 2006])
In [20]: df['data1'].groupby([states, years]).mean()
Out[20]:
California 2005 0.478943
2006 -0.519439
Ohio 2005 -0.380219
2006 1.965781
Name: data1, dtype: float64
In [21]: df.groupby('key1').mean()
Out[21]:
data1 data2
key1
a 0.746672 0.910916
b -0.537585 0.525384
In [22]: df.groupby(['key1', 'key2']).mean()
Out[22]:
data1 data2
key1 key2
a one 0.880536 1.319920
two 0.478943 0.092908
b one -0.519439 0.281746
two -0.555730 0.769023
In [23]: df.groupby(['key1', 'key2']).size()
Out[23]:
key1 key2
a one 2
two 1
b one 1
two 1
dtype: int64
注意,任何分组关键词中的缺失值,都会被从结果中除去。
对分组进行迭代
GroupBy对象支持迭代,可以产生一组二元元组(由分组名和数据块组成)。看下面的例子:
In [24]: for name, group in df.groupby('key1'):
....: print(name)
....: print(group)
....:
a
data1 data2 key1 key2
0 -0.204708 1.393406 a one
1 0.478943 0.092908 a two
4 1.965781 1.246435 a one
b
data1 data2 key1 key2
2 -0.519439 0.281746 b one
3 -0.555730 0.769023 b two
对于多重键的情况,元组的第一个元素将会是由键值组成的元组:
In [25]: for (k1, k2), group in df.groupby(['key1', 'key2']):
....: print((k1, k2))
....: print(group)
....:
('a', 'one')
data1 data2 key1 key2
0 -0.204708 1.393406 a one
4 1.965781 1.246435 a one
('a', 'two')
data1 data2 key1 key2
1 0.478943 0.092908 a two
('b', 'one')
data1 data2 key1 key2
2 -0.519439 0.281746 b one
('b', 'two')
data1 data2 key1 key2
3 -0.55573 0.769023 b two
当然,你可以对这些数据片段做任何操作。有一个你可能会觉得有用的运算:将这些数据片段做成一个字典:
In [26]: pieces = dict(list(df.groupby('key1')))
In [27]: pieces['b']
Out[27]:
data1 data2 key1 key2
2 -0.519439 0.281746 b one
3 -0.555730 0.769023 b two
In [28]: df.dtypes
Out[28]:
data1 float64
data2 float64
key1 object
key2 object
dtype: object
In [29]: grouped = df.groupby(df.dtypes, axis=1)
可以如下打印分组:
In [30]: for dtype, group in grouped:
....: print(dtype)
....: print(group)
....:
float64
data1 data2
0 -0.204708 1.393406
1 0.478943 0.092908
2 -0.519439 0.281746
3 -0.555730 0.769023
4 1.965781 1.246435
object
key1 key2
0 a one
1 a two
2 b one
3 b two
4 a one
In [32]: s_grouped = df.groupby(['key1', 'key2'])['data2']
In [33]: s_grouped
Out[33]: <pandas.core.groupby.SeriesGroupBy object at 0x7faa30c78da0>
In [34]: s_grouped.mean()
Out[34]:
key1 key2
a one 1.319920
two 0.092908
b one 0.281746
two 0.769023
Name: data2, dtype: float64
通过字典或Series进行分组
除数组以外,分组信息还可以其他形式存在。来看另一个示例DataFrame:
In [35]: people = pd.DataFrame(np.random.randn(5, 5),
....: columns=['a', 'b', 'c', 'd', 'e'],
....: index=['Joe', 'Steve', 'Wes', 'Jim', 'Travis'])
In [36]: people.iloc[2:3, [1, 2]] = np.nan # Add a few NA values
In [37]: people
Out[37]:
a b c d e
Joe 1.007189 -1.296221 0.274992 0.228913 1.352917
Steve 0.886429 -2.001637 -0.371843 1.669025 -0.438570
Wes -0.539741 NaN NaN -1.021228 -0.577087
Jim 0.124121 0.302614 0.523772 0.000940 1.343810
Travis -0.713544 -0.831154 -2.370232 -1.860761 -0.860757
In [39]: by_column = people.groupby(mapping, axis=1)
In [40]: by_column.sum()
Out[40]:
blue red
Joe 0.503905 1.063885
Steve 1.297183 -1.553778
Wes -1.021228 -1.116829
Jim 0.524712 1.770545
Travis -4.230992 -2.405455
Series也有同样的功能,它可以被看做一个固定大小的映射:
In [41]: map_series = pd.Series(mapping)
In [42]: map_series
Out[42]:
a red
b red
c blue
d blue
e red
f orange
dtype: object
In [43]: people.groupby(map_series, axis=1).count()
Out[43]:
blue red
Joe 2 3
Steve 2 3
Wes 1 2
Jim 2 3
Travis 2 3
In [44]: people.groupby(len).sum()
Out[44]:
a b c d e
3 0.591569 -0.993608 0.798764 -0.791374 2.119639
5 0.886429 -2.001637 -0.371843 1.669025 -0.438570
6 -0.713544 -0.831154 -2.370232 -1.860761 -0.860757
将函数跟数组、列表、字典、Series混合使用也不是问题,因为任何东西在内部都会被转换为数组:
In [45]: key_list = ['one', 'one', 'one', 'two', 'two']
In [46]: people.groupby([len, key_list]).min()
Out[46]:
a b c d e
3 one -0.539741 -1.296221 0.274992 -1.021228 -0.577087
two 0.124121 0.302614 0.523772 0.000940 1.343810
5 one 0.886429 -2.001637 -0.371843 1.669025 -0.438570
6 two -0.713544 -0.831154 -2.370232 -1.860761 -0.860757
In [51]: df
Out[51]:
data1 data2 key1 key2
0 -0.204708 1.393406 a one
1 0.478943 0.092908 a two
2 -0.519439 0.281746 b one
3 -0.555730 0.769023 b two
4 1.965781 1.246435 a one
In [52]: grouped = df.groupby('key1')
In [53]: grouped['data1'].quantile(0.9)
Out[53]:
key1
a 1.668413
b -0.523068
Name: data1, dtype: float64
如果要使用你自己的聚合函数,只需将其传入aggregate或agg方法即可:
In [54]: def peak_to_peak(arr):
....: return arr.max() - arr.min()
In [55]: grouped.agg(peak_to_peak)
Out[55]:
data1 data2
key1
a 2.170488 1.300498
b 0.036292 0.487276
In [56]: grouped.describe()
Out[56]:
data1 \
count mean std min 25% 50% 75%
key1
a 3.0 0.746672 1.109736 -0.204708 0.137118 0.478943 1.222362
b 2.0 -0.537585 0.025662 -0.555730 -0.546657 -0.537585 -0.528512
data2 \
max count mean std min 25% 50%
key1
a 1.965781 3.0 0.910916 0.712217 0.092908 0.669671 1.246435
b -0.519439 2.0 0.525384 0.344556 0.281746 0.403565 0.525384
75% max
key1
a 1.319920 1.393406
b 0.647203 0.769023
In [57]: tips = pd.read_csv('examples/tips.csv')
# Add tip percentage of total bill
In [58]: tips['tip_pct'] = tips['tip'] / tips['total_bill']
In [59]: tips[:6]
Out[59]:
total_bill tip smoker day time size tip_pct
0 16.99 1.01 No Sun Dinner 2 0.059447
1 10.34 1.66 No Sun Dinner 3 0.160542
2 21.01 3.50 No Sun Dinner 3 0.166587
3 23.68 3.31 No Sun Dinner 2 0.139780
4 24.59 3.61 No Sun Dinner 4 0.146808
5 25.29 4.71 No Sun Dinner 4 0.186240
In [60]: grouped = tips.groupby(['day', 'smoker'])
注意,对于表10-1中的那些描述统计,可以将函数名以字符串的形式传入:
In [61]: grouped_pct = grouped['tip_pct']
In [62]: grouped_pct.agg('mean')
Out[62]:
day smoker
Fri No 0.151650
Yes 0.174783
Sat No 0.158048
Yes 0.147906
Sun No 0.160113
Yes 0.187250
Thur No 0.160298
Yes 0.163863
Name: tip_pct, dtype: float64
如果传入一组函数或函数名,得到的DataFrame的列就会以相应的函数命名:
In [63]: grouped_pct.agg(['mean', 'std', peak_to_peak])
Out[63]:
mean std peak_to_peak
day smoker
Fri No 0.151650 0.028123 0.067349
Yes 0.174783 0.051293 0.159925
Sat No 0.158048 0.039767 0.235193
Yes 0.147906 0.061375 0.290095
Sun No 0.160113 0.042347 0.193226
Yes 0.187250 0.154134 0.644685
Thur No 0.160298 0.038774 0.193350
Yes 0.163863 0.039389 0.151240
In [64]: grouped_pct.agg([('foo', 'mean'), ('bar', np.std)])
Out[64]:
foo bar
day smoker
Fri No 0.151650 0.028123
Yes 0.174783 0.051293
Sat No 0.158048 0.039767
Yes 0.147906 0.061375
Sun No 0.160113 0.042347
Yes 0.187250 0.154134
Thur No 0.160298 0.038774
Yes 0.163863 0.039389
In [68]: result['tip_pct']
Out[68]:
count mean max
day smoker
Fri No 4 0.151650 0.187735
Yes 15 0.174783 0.263480
Sat No 45 0.158048 0.291990
Yes 42 0.147906 0.325733
Sun No 57 0.160113 0.252672
Yes 19 0.187250 0.710345
Thur No 45 0.160298 0.266312
Yes 17 0.163863 0.241255
跟前面一样,这里也可以传入带有自定义名称的一组元组:
In [69]: ftuples = [('Durchschnitt', 'mean'),('Abweichung', np.var)]
In [70]: grouped['tip_pct', 'total_bill'].agg(ftuples)
Out[70]:
tip_pct total_bill
Durchschnitt Abweichung Durchschnitt Abweichung
day smoker
Fri No 0.151650 0.000791 18.420000 25.596333
Yes 0.174783 0.002631 16.813333 82.562438
Sat No 0.158048 0.001581 19.661778 79.908965
Yes 0.147906 0.003767 21.276667 101.387535
Sun No 0.160113 0.001793 20.506667 66.099980
Yes 0.187250 0.023757 24.120000 109.046044
Thur No 0.160298 0.001503 17.113111 59.625081
Yes 0.163863 0.001551 19.190588 69.808518
In [77]: tips.groupby(['smoker', 'day']).apply(top, n=1, column='total_bill')
Out[77]:
total_bill tip smoker day time size tip_pct
smoker day
No Fri 94 22.75 3.25 No Fri Dinner 2 0.142857
Sat 212 48.33 9.00 No Sat Dinner 4 0.186220
Sun 156 48.17 5.00 No Sun Dinner 6 0.103799
Thur 142 41.19 5.00 No Thur Lunch 5 0.121389
Yes Fri 95 40.17 4.73 Yes Fri Dinner 4 0.117750
Sat 170 50.81 10.00 Yes Sat Dinner 3 0.196812
Sun 182 45.35 3.50 Yes Sun Dinner 3 0.077178
Thur 197 43.11 5.00 Yes Thur Lunch 4 0.115982
In [78]: result = tips.groupby('smoker')['tip_pct'].describe()
In [79]: result
Out[79]:
count mean std min 25% 50% 75% \
smoker
No 151.0 0.159328 0.039910 0.056797 0.136906 0.155625 0.185014
Yes 93.0 0.163196 0.085119 0.035638 0.106771 0.153846 0.195059
max
smoker
No 0.291990
Yes 0.710345
In [80]: result.unstack('smoker')
Out[80]:
smoker
count No 151.000000
Yes 93.000000
mean No 0.159328
Yes 0.163196
std No 0.039910
Yes 0.085119
min No 0.056797
Yes 0.035638
25% No 0.136906
Yes 0.106771
50% No 0.155625
Yes 0.153846
75% No 0.185014
Yes 0.195059
max No 0.291990
Yes 0.710345
dtype: float64
In [99]: data[['Vermont', 'Nevada', 'Idaho']] = np.nan
In [100]: data
Out[100]:
Ohio 0.922264
New York -2.153545
Vermont NaN
Florida -0.375842
Oregon 0.329939
Nevada NaN
California 1.105913
Idaho NaN
dtype: float64
In [101]: data.groupby(group_key).mean()
Out[101]:
East -0.535707
West 0.717926
dtype: float64
我们可以用分组平均值去填充NA值:
In [102]: fill_mean = lambda g: g.fillna(g.mean())
In [103]: data.groupby(group_key).apply(fill_mean)
Out[103]:
Ohio 0.922264
New York -2.153545
Vermont -0.535707
Florida -0.375842
Oregon 0.329939
Nevada 0.717926
California 1.105913
Idaho 0.717926
dtype: float64
另外,也可以在代码中预定义各组的填充值。由于分组具有一个name属性,所以我们可以拿来用一下:
In [104]: fill_values = {'East': 0.5, 'West': -1}
In [105]: fill_func = lambda g: g.fillna(fill_values[g.name])
In [106]: data.groupby(group_key).apply(fill_func)
Out[106]:
Ohio 0.922264
New York -2.153545
Vermont 0.500000
Florida -0.375842
Oregon 0.329939
Nevada -1.000000
California 1.105913
Idaho -1.000000
dtype: float64
示例:随机采样和排列
假设你想要从一个大数据集中随机抽取(进行替换或不替换)样本以进行蒙特卡罗模拟(Monte Carlo simulation)或其他分析工作。“抽取”的方式有很多,这里使用的方法是对Series使用sample方法:
In [111]: get_suit = lambda card: card[-1] # last letter is suit
In [112]: deck.groupby(get_suit).apply(draw, n=2)
Out[112]:
C 2C 2
3C 3
D KD 10
8D 8
H KH 10
3H 3
S 2S 2
4S 4
dtype: int64
In [114]: df = pd.DataFrame({'category': ['a', 'a', 'a', 'a',
.....: 'b', 'b', 'b', 'b'],
.....: 'data': np.random.randn(8),
.....: 'weights': np.random.rand(8)})
In [115]: df
Out[115]:
category data weights
0 a 1.561587 0.957515
1 a 1.219984 0.347267
2 a -0.482239 0.581362
3 a 0.315667 0.217091
4 b -0.047852 0.894406
5 b -0.454145 0.918564
6 b -0.556774 0.277825
7 b 0.253321 0.955905
然后可以利用category计算分组加权平均数:
In [116]: grouped = df.groupby('category')
In [117]: get_wavg = lambda g: np.average(g['data'], weights=g['weights'])
In [118]: grouped.apply(get_wavg)
Out[118]:
category
a 0.811643
b -0.122262
dtype: float64
顺着上一个例子继续,你可以用groupby执行更为复杂的分组统计分析,只要函数返回的是pandas对象或标量值即可。例如,我可以定义下面这个regress函数(利用statsmodels计量经济学库)对各数据块执行普通最小二乘法(Ordinary Least Squares,OLS)回归:
import statsmodels.api as sm
def regress(data, yvar, xvars):
Y = data[yvar]
X = data[xvars]
X['intercept'] = 1.
result = sm.OLS(Y, X).fit()
return result.params
In [133]: tips.pivot_table('tip_pct', index=['time', 'smoker'], columns='day',
.....: aggfunc=len, margins=True)
Out[133]:
day Fri Sat Sun Thur All
time smoker
Dinner No 3.0 45.0 57.0 1.0 106.0
Yes 9.0 42.0 19.0 NaN 70.0
Lunch No 1.0 NaN NaN 44.0 45.0
Yes 6.0 NaN NaN 17.0 23.0
All 19.0 87.0 76.0 62.0 244.0
如果存在空的组合(也就是NA),你可能会希望设置一个fill_value:
In [134]: tips.pivot_table('tip_pct', index=['time', 'size', 'smoker'],
.....: columns='day', aggfunc='mean', fill_value=0)
Out[134]:
day Fri Sat Sun Thur
time size smoker
Dinner 1 No 0.000000 0.137931 0.000000 0.000000
Yes 0.000000 0.325733 0.000000 0.000000
2 No 0.139622 0.162705 0.168859 0.159744
Yes 0.171297 0.148668 0.207893 0.000000
3 No 0.000000 0.154661 0.152663 0.000000
Yes 0.000000 0.144995 0.152660 0.000000
4 No 0.000000 0.150096 0.148143 0.000000
Yes 0.117750 0.124515 0.193370 0.000000
5 No 0.000000 0.000000 0.206928 0.000000
Yes 0.000000 0.106572 0.065660 0.000000
... ... ... ... ...
Lunch 1 No 0.000000 0.000000 0.000000 0.181728
Yes 0.223776 0.000000 0.000000 0.000000
2 No 0.000000 0.000000 0.000000 0.166005
Yes 0.181969 0.000000 0.000000 0.158843
3 No 0.187735 0.000000 0.000000 0.084246
Yes 0.000000 0.000000 0.000000 0.204952
4 No 0.000000 0.000000 0.000000 0.138919
Yes 0.000000 0.000000 0.000000 0.155410
5 No 0.000000 0.000000 0.000000 0.121389
6 No 0.000000 0.000000 0.000000 0.173706
[21 rows x 4 columns]
In [138]: data
Out[138]:
Sample Nationality Handedness
0 1 USA Right-handed
1 2 Japan Left-handed
2 3 USA Right-handed
3 4 Japan Right-handed
4 5 Japan Left-handed
5 6 Japan Right-handed
6 7 USA Right-handed
7 8 USA Left-handed
8 9 Japan Right-handed
9 10 USA Right-handed
In [139]: pd.crosstab(data.Nationality, data.Handedness, margins=True)
Out[139]:
Handedness Left-handed Right-handed All
Nationality
Japan 2 3 5
USA 1 4 5
All 3 7 10
crosstab的前两个参数可以是数组或Series,或是数组列表。就像小费数据:
In [140]: pd.crosstab([tips.time, tips.day], tips.smoker, margins=True)
Out[140]:
smoker No Yes All
time day
Dinner Fri 3 9 12
Sat 45 42 87
Sun 57 19 76
Thur 1 0 1
Lunch Fri 1 6 7
Thur 44 17 61
All 151 93 244