ENH: DataFrameGroupby.value_counts

[corriebar]

• I have checked that this issue has not already been reported.

• I have confirmed this bug exists on the latest version of pandas.

• I have confirmed this bug exists on the master branch of pandas.

Reproducible Example

import pandas as pd

df = pd.DataFrame({'gender': ['female', 'male', 'female', 'male', 'female', 'male'], 
                  'education': ['low', 'medium', 'high', 'low', 'high', 'low'],
                  'country': ['US', 'FR', 'US', 'FR', 'FR', 'US']})
# These all work fine:
# value_counts() on a DataFrame
df[['gender', 'education']].value_counts(normalize=True)
df['gender'].value_counts(normalize=True) # on a Series
df.groupby('country')['gender'].value_counts(normalize=True) # on a SeriesGroupby

# but fails on DataFrameGroupBy
df.groupby('country')[['gender', 'education']].value_counts(normalize=True)
---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
/var/folders/xz/_nt7tl1s0jx2f5_8rwc8kc0r0000gp/T/ipykernel_96325/3153685003.py in <module>
----> 1 df.groupby('country')[['gender', 'education']].value_counts(normalize=True)

~/opt/anaconda3/envs/pandas_dev/lib/python3.9/site-packages/pandas/core/groupby/groupby.py in __getattr__(self, attr)
    909             return self[attr]
    910 
--> 911         raise AttributeError(
    912             f"'{type(self).__name__}' object has no attribute '{attr}'"
    913         )

AttributeError: 'DataFrameGroupBy' object has no attribute 'value_counts'

Issue Description

Since value_counts() is defined both for a Series and a DataFrame, I also expect it to work on both a SeriesGroupBy and a DataFrameGroupBy.

There are some related Issues: #39938 and #6540, these have been dismissed so far with the argument that size() already does that, but size() alone is not enough to get the proportions per group.

Expected Behavior

I managed to get the desired behaviour by applying value_counts() to each group:

def compute_proportions(df, var):
    return (df[var]
        .value_counts(normalize=True)
    )
(df.groupby('country')
  .apply(compute_proportions, var=['gender', 'education'])
)
country  gender  education
FR       female  high         0.333333
         male    low          0.333333
                 medium       0.333333
US       male    low          0.666667
         female  high         0.333333
dtype: float64

Installed Versions

INSTALLED VERSIONS

commit : 73c6825
python : 3.9.7.final.0
python-bits : 64
OS : Darwin
OS-release : 19.6.0
Version : Darwin Kernel Version 19.6.0: Tue Jun 22 19:49:55 PDT 2021; root:xnu-6153.141.35~1/RELEASE_X86_64
machine : x86_64
processor : i386
byteorder : little
LC_ALL : None
LANG : None
LOCALE : None.UTF-8

pandas : 1.3.3
numpy : 1.21.2
pytz : 2021.1
dateutil : 2.8.2
pip : 21.2.4
setuptools : 58.0.4
Cython : None
pytest : None
hypothesis : None
sphinx : None
blosc : None
feather : None
xlsxwriter : None
lxml.etree : None
html5lib : None
pymysql : None
psycopg2 : None
jinja2 : None
IPython : 7.27.0
pandas_datareader: None
bs4 : None
bottleneck : None
fsspec : None
fastparquet : None
gcsfs : None
matplotlib : None
numexpr : None
odfpy : None
openpyxl : None
pandas_gbq : None
pyarrow : None
pyxlsb : None
s3fs : None
scipy : None
sqlalchemy : None
tables : None
tabulate : None
xarray : None
xlrd : None
xlwt : None
numba : None

[MarcoGorelli]

Hi @corriebar

I think this is a good request, thanks for including clear input + output + workaround

@rhshadrach any thoughts on this?

[rhshadrach]

Thanks for reporting this, and agree with @MarcoGorelli - well written.

There are some related Issues: #39938 and #6540, these have been dismissed so far with the argument that size() already does that, but size() alone is not enough to get the proportions per group.

#6540 was asking for counts of the grouping columns, rather than the rest of the columns in the DataFrame. This is very much different from the request here. For #39938, while size was mentioned in the comments, this issue was closed because it lacked a reproducible example demonstrating the request.

In general I agree with having a consistent API between Series / DataFrame/ SeriesGroupBy / DataFrameGroupby when the operation makes sense, and that seems to me to be the case here. The oddity is that value_counts doesn't fit into one of the agg / transform / filter buckets that most groupby ops do.

The index should be the groups for agg, the original index for transform, and a subset of the original index for a filter - but because value_counts doesn't fit into one of these, I think it's okay that the resulting index be the grouping columns combined with the other columns in the DataFrame. This then agrees with Series.value_counts, DataFrame.value_counts and SeriesGroupBy.value_counts, so the proposed result looks good to me.

Looking at the code for SeriesGroupBy.value_counts, it seems like an implementation for DataFrameGroupBy would be non-trivial. Here is a naive attempt to use size that seems to perform well when compared to the SeriesGroupBy variant, but I'm guessing it will fail on various edge cases.

def gb_value_counts(df, keys, subset=None, normalize=False, sort=True, ascending=False, dropna=True):
    remaining_columns = list(df.columns.difference(keys))
    if subset is not None:
        remaining_columns = list(set(subset) - set(remaining_columns))
    if dropna:
        df = df.dropna(subset=remaining_columns, axis='index', how='any')
    result = df.groupby(keys + remaining_columns).size()
    if normalize:
        result /= df.groupby(keys).size()
    if sort:
        result = result.sort_values(ascending=ascending)
    return result

Timing:

size = 100000
df = pd.DataFrame(
    {
        'a': np.random.randint(0, 10, size),
        'b': np.random.randint(0, 10, size),
    }
).set_index(['a'])

%timeit gb_value_counts(df, ['a'], normalize=False, dropna=True)  # 5.83 ms ± 26.8 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)
%timeit df['b'].groupby('a').value_counts()  # 9.44 ms ± 44.3 µs per loop (mean ± std. dev. of 7 runs, 100 loops each)

[johnzangwill]

take

[johnzangwill]

@rhshadrach I have implemented #44267 loosely based on your suggestion.
Please could you take a look at the implementation and tests to see if there are any cases that I have missed.
If possible I would like this in 1.4, so I need to start getting it reviewed...
Thanks!