Aggregate mean over TimeDelta column

[AlbertDeFusco]

Cannot compute mean of timedelta in gropuby

In [1]: import pandas as pd
   ...: import numpy as np
   ...: print(pd.__version__)
   ...: 
   ...: time = pd.to_timedelta(np.random.randint(100,1000, size=10), unit='s')
   ...: cat = np.random.choice(['A','B','C'], size=10)
   ...: 
   ...: df = pd.DataFrame(dict(time=time, cat=cat))
   ...: 
   ...: print(df['time'].mean())
   ...: 
   ...: df.groupby('cat')['time'].mean()
   ...: 
0.20.3
0 days 00:07:15.800000
---------------------------------------------------------------------------
DataError                                 Traceback (most recent call last)
<ipython-input-1-b649ebd78333> in <module>()
     10 print(df['time'].mean())
     11 
---> 12 df.groupby('cat')['time'].mean()

/Users/adefusco/Applications/anaconda3/4.3/envs/module-pandas/lib/python3.6/site-packages/pandas/core/groupby.py in mean(self, *args, **kwargs)
   1035         nv.validate_groupby_func('mean', args, kwargs, ['numeric_only'])
   1036         try:
-> 1037             return self._cython_agg_general('mean', **kwargs)
   1038         except GroupByError:
   1039             raise

/Users/adefusco/Applications/anaconda3/4.3/envs/module-pandas/lib/python3.6/site-packages/pandas/core/groupby.py in _cython_agg_general(self, how, alt, numeric_only)
    834 
    835         if len(output) == 0:
--> 836             raise DataError('No numeric types to aggregate')
    837 
    838         return self._wrap_aggregated_output(output, names)

DataError: No numeric types to aggregate

Problem description

The mean of TimeDelta can be computed, but it is not working in GroupBy.

Expected Output

This is the output I would expect.

In [3]: def average_time(x):
   ...:     s = x.dt.seconds
   ...:     return pd.Timedelta(s.mean(), unit='s')
   ...: 
   ...: df.groupby('cat')['time'].apply(average_time)
   ...: 
Out[3]: 
cat
A   00:09:27.666667
B   00:06:18.666667
C   00:06:19.750000
Name: time, dtype: timedelta64[ns]

Output of pd.show_versions()

INSTALLED VERSIONS

commit: None
python: 3.6.1.final.0
python-bits: 64
OS: Darwin
OS-release: 16.7.0
machine: x86_64
processor: i386
byteorder: little
LC_ALL: None
LANG: en_US.UTF-8
LOCALE: en_US.UTF-8

pandas: 0.20.3
pytest: None
pip: 9.0.1
setuptools: 27.2.0
Cython: None
numpy: 1.13.1
scipy: 0.19.1
xarray: None
IPython: 6.0.0
sphinx: None
patsy: 0.4.1
dateutil: 2.6.0
pytz: 2017.2
blosc: None
bottleneck: None
tables: 3.4.2
numexpr: 2.6.2
feather: None
matplotlib: 2.0.2
openpyxl: 2.4.1
xlrd: 1.0.0
xlwt: 1.2.0
xlsxwriter: None
lxml: None
bs4: None
html5lib: 0.999
sqlalchemy: 1.1.9
pymysql: None
psycopg2: None
jinja2: 2.9.6
s3fs: None
pandas_gbq: None
pandas_datareader: 0.2.1

[TomAugspurger]

I think this is just because we don't have a cython implementation of group_mean in

pandas/pandas/_libs/groupby_helper.pxi.in

Line 31 in ad7d6fc

{{for name, c_type, dest_type, dest_type2, dest_dtype in get_dispatch(dtypes)}}

if you're interested in taking a look

FYI, a slightly smaller workaround is df.groupby('cat')['time'].agg(lambda x: x.mean())

[TomAugspurger]

Oh, it might be a bit more involved than that, since it looks like we call ._get_numeric_data, which excludes timedeltas.

[chris-b1]

The cython path will handle timedeltas as below

In [47]: df.groupby('cat')['time'].mean(numeric_only=False)
Out[47]: 
cat
A   00:09:46.666666
B   00:10:02.750000
Name: time, dtype: timedelta64[ns]

Arguably not a great API - we automatically exclude "non-numeric" columns on groupby ops. This makes sense with an entire DataFrame, but perhaps on a Series selection this should default to true?

[Holt59]

@chris-b1 Just tried this on my dataframe, and it does not give me correct results, I think it's because it handles NaT incorrectly (it gives me negative Timedelta from a dataframe containing only positive Timedelta and NaT).

[MattW216]

This same issue is occurring when using .resample(). I'm guessing that uses GroupBy in some fashion, but in case it does not, I thought I'd mention it here.

Also, @chris-b1 's workaround worked just fine for me.

[qinyugwu]

It seems to make sense to convert timestamp column to either days or seconds by using the inbuilt functions to get the mean.

[chbrown]

FWIW (another workaround), you can use ✅ ...agg({'duration': pd.Series.mean}) instead of ❌ ...agg({'duration': 'mean'}) or ❌ ...agg({'duration': np.mean}) (where duration is a timedelta64-typed column).