Decile calculation with "ntile"...

[Carlos Ortega]

Hi,

Thanks for your excellent package to port R (dplyr) flow of processing to Python. I have been using another alternatives, and yours is the one that offers the most extensive and equivalent to what is possible now with dplyr.

I have an issue with how ntile() calculates the different groups for a vector of probabilities ("p2").

This is the output of that calculation.

kk = ( trainnonFun >> select( f.p2) >> mutate( decil = ntile(f.p2, n=10)))
kk
            p2      decil
      <float64> <category>
6535   0.971462         10
7523   0.971462         10
48441  0.970154         10
48417  0.970154         10
...         ...        ...
13971  0.970154         10
38140  0.409739          1
13400  0.409739          1
45999  0.405575          1
26150  0.372226          1
29939  0.357850          1

But when you calculates how many values are in each bucket, it shows something strange:

pp = ( kk >> count(f.decil))
pp
      decil       n
  <category> <int64>
0          1       7
1          2     542
2          3    1361
3          4     924
4          5    1240
5          6    1655
6          7    3080
7          8    2647
8          9    1571
9         10    1345

The groups are very dissimilar.

For the sake of reproducibility, In this file you can find that dataframe with the probabilities and the calculated decile.

• https://github.com/coforfe/deciles/blob/main/deciles.csv

Now, I am calculating the right decile with pandas qcut() method, which offers the right output, with a much mofre balanced number of elements in each bucket.

Thanks again,
Carlos.

[pwwang]

Hi, could you show the versions of datar:

from datar import get_versions
get_versions()

[Carlos Ortega]

Hi,
Yes, this is what I get.

>>> from datar import get_versions
>>> get_versions()
python      : 3.10.4 (main, Mar 31 2022, 08:41:55) [GCC 7.5.0]
datar       : 0.10.2
simplug     : 0.2.1
executing   : 1.2.0
pipda       : 0.10.0
datar-numpy : 0.0.0
numpy       : 1.23.4
datar-pandas: 0.0.0
pandas      : 1.5.2

Thanks!
Carlos.

[pwwang]

This is a nice catch!
The ntile implementation should use pd.qcut instead of pd.cut.
It shall be fixed by datar-pandas v0.1.1

Try updating datar by:

pip install -U datar[pandas]

and also try get_versions to ensure datar-pandas v0.1.1 is installed.

[pwwang]

By the way, thanks for the compliments:

Thanks for your excellent package to port R (dplyr) flow of processing to Python. I have been using other alternatives, and yours is the one that offers the most extensive and equivalent to what is possible now with dplyr.

Do you mind if I put it as a testimonial in the README file?

[Carlos Ortega]

Thanks a lot for your quick fix!.

No, I do not mind at all.
Thanks to you.
Carlos.

[pwwang]

Thanks!

Please confirm if this is fixed and feel free to close it if so.
Feel free to open new issues if you have other questions.

[Carlos Ortega]

Thanks,

Yes, I have just updated datar with your indications and now the problem is fixed.

       decil       n
  <category> <int64>
0          1    1438
1          2    1437
2          3    1437
3          4    1437
4          5    1437
5          6    1439
6          7    1435
7          8    1585
8          9    1293
9         10    1434

Thanks again,
Carlos.