Welcome to OGeek Q&A Community for programmer and developer-Open, Learning and Share
Welcome To Ask or Share your Answers For Others

Categories

0 votes
293 views
in Technique[技术] by (71.8m points)

python - Winsorizing data by column in pandas with NaN

I'd like to winsorize several columns of data in a pandas Data Frame. Each column has some NaN, which affects the winsorization, so they need to be removed. The only way I know how to do this is to remove them for all of the data, rather than remove them only column-by-column.

MWE:

import numpy as np
import pandas as pd
from scipy.stats.mstats import winsorize

# Create Dataframe
N, M, P = 10**5, 4, 10**2
dates = pd.date_range('2001-01-01', periods=N//P, freq='D').repeat(P)
df = pd.DataFrame(np.random.random((N, M))
                  , index=dates)
df.index.names = ['DATE']
df.columns = ['one','two','three','four']
# Now scale them differently so you can see the winsorization
df['four'] = df['four']*(10**5)
df['three'] = df['three']*(10**2)
df['two'] = df['two']*(10**-1)
df['one'] = df['one']*(10**-4)
# Create NaN
df.loc[df.index.get_level_values(0).year == 2002,'three'] = np.nan
df.loc[df.index.get_level_values(0).month == 2,'two'] = np.nan
df.loc[df.index.get_level_values(0).month == 1,'one'] = np.nan

Here is the baseline distribution:

df.quantile([0, 0.01, 0.5, 0.99, 1])

output:

               one           two      three          four
0.00  2.336618e-10  2.294259e-07   0.002437      2.305353
0.01  9.862626e-07  9.742568e-04   0.975807   1003.814520
0.50  4.975859e-05  4.981049e-02  50.290946  50374.548980
0.99  9.897463e-05  9.898590e-02  98.978263  98991.438985
1.00  9.999983e-05  9.999966e-02  99.996793  99999.437779

This is how I'm winsorizing:

def using_mstats(s):
    return winsorize(s, limits=[0.01, 0.01])

wins = df.apply(using_mstats, axis=0)
wins.quantile([0, 0.01, 0.25, 0.5, 0.75, 0.99, 1])

Which gives this:

Out[356]:
           one       two      three          four
0.00  0.000001  0.001060   1.536882   1003.820149
0.01  0.000001  0.001060   1.536882   1003.820149
0.25  0.000025  0.024975  25.200378  25099.994780
0.50  0.000050  0.049810  50.290946  50374.548980
0.75  0.000075  0.074842  74.794537  75217.343920
0.99  0.000099  0.098986  98.978263  98991.436957
1.00  0.000100  0.100000  99.996793  98991.436957

Column four is correct because it has no NaN but the others are incorrect. The 99th percentile and Max should be the same. The observations counts are identical for both:

In [357]: df.count()
Out[357]:
one       90700
two       91600
three     63500
four     100000
dtype: int64

In [358]: wins.count()
Out[358]:
one       90700
two       91600
three     63500
four     100000
dtype: int64

This is how I can 'solve' it, but at the cost of losing a lot of my data:

wins2 = df.loc[df.notnull().all(axis=1)].apply(using_mstats, axis=0)
wins2.quantile([0, 0.01, 0.25, 0.5, 0.75, 0.99, 1])

Output:

Out[360]:
               one       two      three          four
0.00  9.686203e-07  0.000928   0.965702   1005.209503
0.01  9.686203e-07  0.000928   0.965702   1005.209503
0.25  2.486052e-05  0.024829  25.204032  25210.837443
0.50  4.980946e-05  0.049894  50.299004  50622.227179
0.75  7.492750e-05  0.075059  74.837900  75299.906415
0.99  9.895563e-05  0.099014  98.972310  99014.311761
1.00  9.895563e-05  0.099014  98.972310  99014.311761

In [361]: wins2.count()
Out[361]:
one      51700
two      51700
three    51700
four     51700
dtype: int64

How can I winsorize the data, by column, that is not NaN, while maintaining the data shape (i.e. not removing rows)?

See Question&Answers more detail:os

与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
Welcome To Ask or Share your Answers For Others

1 Reply

0 votes
by (71.8m points)

As often happens, simply creating the MWE helped clarify. I need to use clip() in combination with quantile() as below:

df2 = df.clip(lower=df.quantile(0.01), upper=df.quantile(0.99), axis=1)
df2.quantile([0, 0.01, 0.25, 0.5, 0.75, 0.99, 1])

Output:

               one       two      three          four
0.00  9.862626e-07  0.000974   0.975807   1003.814520
0.01  9.862666e-07  0.000974   0.975816   1003.820092
0.25  2.485043e-05  0.024975  25.200378  25099.994780
0.50  4.975859e-05  0.049810  50.290946  50374.548980
0.75  7.486737e-05  0.074842  74.794537  75217.343920
0.99  9.897462e-05  0.098986  98.978245  98991.436977
1.00  9.897463e-05  0.098986  98.978263  98991.438985

In [384]: df2.count()
Out[384]:
one       90700
two       91600
three     63500
four     100000
dtype: int64

The numbers are different from above because I have maintained all of the data in each column that is not missing (NaN).


与恶龙缠斗过久,自身亦成为恶龙;凝视深渊过久,深渊将回以凝视…
OGeek|极客中国-欢迎来到极客的世界,一个免费开放的程序员编程交流平台!开放,进步,分享!让技术改变生活,让极客改变未来! Welcome to OGeek Q&A Community for programmer and developer-Open, Learning and Share
Click Here to Ask a Question

...