pandas.cut: the 'include_lowest' argument isn't behaving as documented

[evarga]

Code Sample

import pandas as pd
import numpy as np

pd.cut(np.array([1, 7, 5, 4, 6, 3]), bins=[0, 3, 6, 8], include_lowest=True)

Problem description

Just by setting the include_lowest to True the data type of the interval changes from int64 to float64 and the first interval isn't left-inclusive. Here is the wrong output that you'll get:

[(-0.001, 3.0], (6.0, 8.0], (3.0, 6.0], (3.0, 6.0], (3.0, 6.0], (-0.001, 3.0]]
Categories (3, interval[float64]): [(-0.001, 3.0] < (3.0, 6.0] < (6.0, 8.0]]

Expected Output

[(0, 3], (6, 8], (3, 6], (3, 6], (3, 6], (0, 3]]
Categories (3, interval[int64]): [[0, 3] < (3, 6] < (6, 8]]

Output of pd.show_versions()

INSTALLED VERSIONS ------------------ commit: None python: 3.6.6.final.0 python-bits: 64 OS: Darwin OS-release: 17.7.0 machine: x86_64 processor: i386 byteorder: little LC_ALL: en_US.UTF-8 LANG: en_US.UTF-8 LOCALE: en_US.UTF-8

pandas: 0.23.4
pytest: 3.8.2
pip: 10.0.1
setuptools: 40.4.3
Cython: 0.28.5
numpy: 1.15.2
scipy: 1.1.0
pyarrow: None
xarray: None
IPython: 7.0.1
sphinx: 1.8.1
patsy: 0.5.0
dateutil: 2.7.3
pytz: 2018.5
blosc: None
bottleneck: 1.2.1
tables: 3.4.4
numexpr: 2.6.8
feather: None
matplotlib: 3.0.0
openpyxl: 2.5.8
xlrd: 1.1.0
xlwt: 1.2.0
xlsxwriter: 1.1.1
lxml: 4.2.5
bs4: 4.6.3
html5lib: 1.0.1
sqlalchemy: 1.2.12
pymysql: None
psycopg2: None
jinja2: 2.10
s3fs: None
fastparquet: None
pandas_gbq: None
pandas_datareader: None

[WWH98932]

I met the same problem, I changed the intervals to str format and fix the first bound manually.

[mroeschke]

I think this is the documented behavior as described in the bins section of the docstring:

The criteria to bin by.
int : Defines the number of equal-width bins in the range of x. The range of x is extended by .1% on each side to include the minimum and maximum values of x.

This situation could be better documented though

[evarga]

Thanks for the comment! Please, also see the issue with altered data type of intervals (see my oiginal comment above). This behavior is nowhere documented as far as I know, and could be the consequence of this %1 extension, as you have cited.

[Nakai-Naoto]

The criteria to bin by.
* int : Defines the number of equal-width bins in the range of x. The range of x is extended by .1% on each side to include the minimum and maximum values of x.
* sequence of scalars : Defines the bin edges allowing for non-uniform width. No extension of the range of x is done.

This does not seem to be documented because bins is a sequence of scalars in the code sample.

[gdex1]

I've looked more into how to resolve this and it seems this is more a restriction of IntervalIndex than a documentation error. IntervalIndex requires that all bins are closed on the same side so
(3, interval[int64]): [[0, 3] < (3, 6] < (6, 8]] is not valid.

This is solved by slightly reducing the lower bound of the first interval and making it lower exclusive. Unfortunately, this converts the interval from int64 to float64 and creates the following unexpected behavior:

In: 
pd.cut(np.array([-0.0001, 0, 1, 7, 5, 4, 6, 3, 8]), bins=[0, 3, 6, 8], include_lowest=True)

Out: 
[NaN, (-0.001, 3.0], (-0.001, 3.0], (6.0, 8.0], (3.0, 6.0], (3.0, 6.0], (3.0, 6.0], (-0.001, 3.0], (6.0, 8.0]]
Categories (3, interval[float64]): [(-0.001, 3.0] < (3.0, 6.0] < (6.0, 8.0]]

You can see that even though -0.0001 should be in the first interval, it is not assigned to the first bin.

I am new to this so I was wondering what the best way to handle this is?

[jotasi]

As suggested in #42212 (@mroeschke suggested to include it into the discussion here instead), I would love to also have the option to include not only the left-most boundary for left-open intervals but also the right-most boundary for right-open intervals (i.e. for right=False).

My preferred solution would be to change include_lowest to something like final_interval_closed, which is also functional for right-open intervals (i.e. if right=False is specified).

Maybe, if the solution to this issue is to adapt the function or the underlying IntervalIndex instead of updating the documentation, this could be kept in mind and implemented as well.

[edwhuang23]

take