DataFrame.nlargest result error

[flystarhe]

Code Sample, a copy-pastable example if possible

import pandas as pd
import numpy as np

df = pd.DataFrame({'a': [1, 10, 8, 11, 8],
    'b': list('abdce'),
    'c': [1.0, 2.0, np.nan, 3.0, 4.0]})
print('_________')
print(df.nlargest(10,['a','b']))

Problem description

DataFrame的nlargest在遇到rank相同的情况时，结果错误。如下，第二行和第四行反复出现了。

Expected Output

    a  b    c
3  11  c  3.0
1  10  b  2.0
2   8  d  NaN
4   8  e  4.0
2   8  d  NaN
4   8  e  4.0
0   1  a  1.0
[Finished in 0.6s]

Output of pd.show_versions()

INSTALLED VERSIONS
------------------
commit: None
python: 3.6.0.final.0
python-bits: 64
OS: Windows
OS-release: 10
machine: AMD64
processor: Intel64 Family 6 Model 94 Stepping 3, GenuineIntel
byteorder: little
LC_ALL: None
LANG: None
LOCALE: None.None

[jreback]

what exactly is the problem? show the full pd.show_versions() as well.

[jreback]

This is exactly what should be reported.

In [9]: df.nlargest(10, columns=['a', 'b'])
TypeError: Column 'b' has dtype object, cannot use method 'nlargest' with this dtype

[oldclesleycode]

Yeah, so it appears that it doesn't handle dtypes that are "objects" since it wasn't built to handle sorting strings.

>>> df.dtypes
a      int64
b     object
c    float64

[flystarhe]

@jreback @lesley2958
no, gourp by key list exist duplicates:

tmp = df.nlargest(10,['a','b']).index.unique()
print(df.loc[tmp])

output:

    a  b    c
3  11  c  3.0
1  10  b  2.0
2   8  d  NaN
4   8  e  4.0
0   1  a  1.0

[jorisvandenbossche]

What might be a reason for confusion here (depending on the version @flystarhe is using) is that there is a difference between 0.19.2 and 0.20

In [6]: df.nlargest(10,['a','b'])
Out[6]: 
    a  b    c
3  11  c  3.0
1  10  b  2.0
2   8  d  NaN
4   8  e  4.0
2   8  d  NaN
4   8  e  4.0
0   1  a  1.0

In [7]: pd.__version__
Out[7]: '0.19.2'

In [57]: df.nlargest(10,['a','b'])
...
TypeError: Column 'b' has dtype object, cannot use method 'nlargest' with this dtype

In [58]: pd.__version__
Out[58]: '0.21.0.dev+19.g69a5d6f.dirty'

That said, the output of 0.19.2 also seems wrong (even if object columns would be allowed). -> but that seems to be fixed: #15297

[jorisvandenbossche]

Given that the following methods that rely on order work for object dtype:

In [62]: pd.Series(['a', 'b', 'd', 'c']).sort_values()
Out[62]: 
0    a
1    b
3    c
2    d
dtype: object

In [63]: pd.Series(['a', 'b', 'd', 'c']).max()
Out[63]: 'd'

you could also say nlargest should work for object dtype.

[flystarhe]

@jorisvandenbossche But it doesn't solve the problem. Because nlargest has different efficiencies and computational idea

[jreback]

This already raised for Series in 0.19.2, but not for DataFrame. This was unified to disallow object columns generally and take on the Series behavior (and of course fix the actual duplicated issues)
xref #15299

In [3]: Series(list('abc')).nlargest(1)
TypeError: Cannot use method 'nlargest' with dtype object

[jreback]

@flystarhe your question is not clear

These might be what you want

In [5]: df.groupby(['a', 'b']).nth([0, 1, 2])
Out[5]: 
        c
a  b     
1  a  1.0
8  d  NaN
   e  4.0
10 b  2.0
11 c  3.0

In [6]: df.sort_values(['a', 'b']).groupby(['a', 'b']).head(10)
Out[6]: 
    a  b    c
0   1  a  1.0
2   8  d  NaN
4   8  e  4.0
1  10  b  2.0
3  11  c  3.0

[jorisvandenbossche]

But it doesn't solve the problem. Because nlargest has different efficiencies and computational idea

It was not meant to solve your problem (which you should try to explain better). I was just giving a possible reason to allow nlargest on object columns. But since we also raise for series, I don't think we are going to change this.

[mroeschke]

Closing as the current behavior is intended and correct.