Columns lose category dtype after calling replace on the dataframe

[somiandras]

xref #25521

Code Sample, a copy-pastable example if possible

>>> import pandas as pd
>>> pd.__version__
'0.23.4'

# Sample dataframe with two categorical and one int column
>>> df = pd.DataFrame(
...     [[1., 'A', 'x'], [2, 'B', 'y'], [3, 'C', 'z']],
...     columns=['first', 'second', 'third']
... ).astype({'second': 'category', 'third': 'category'})

# Replace int values
>>> df = df.replace(1, 10)

# Both categorical columns turned into object...
>>> df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 3 columns):
first     3 non-null float64
second    3 non-null object
third     3 non-null object
dtypes: float64(1), object(2)
memory usage: 152.0+ bytes

# Sample dataframe with two categorical columns
>>> df = pd.DataFrame(
...     [['A', 'x'], ['B', 'y'], ['C', 'z']],
...     columns=['first', 'second',]
... ).astype({'first': 'category', 'second': 'category'})

>>> df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 2 columns):
first     3 non-null category
second    3 non-null category
dtypes: category(2)
memory usage: 294.0 bytes

# Replace values in column `first`
>>> df = df.replace('A', 'B')

# Dtype of column `second` becomes object
>>> df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 2 columns):
first     3 non-null category
second    3 non-null object
dtypes: category(1), object(1)
memory usage: 211.0+ bytes

Problem description

Calling replace() on a dataframe changes category columns' dtype to object in an apparently inconsisent manner:

1. When replace() contains non-categorical values that are not present in either categorical column, all categorical dtype turn into object.
2. When replace() contains categories from a category dtype column, then that column keeps its dtype, but other categorical columns turn into object.

See the examples above for both. Even if it is somehow intentional I find it quite confusing.

Expected Output

I would expect the categorical columns to keep their category dtype after replace is called on the dataframe (at least for those categorical columns that are unaffected by replace).

Output of pd.show_versions()

INSTALLED VERSIONS

commit: None

pandas: 0.23.4
pytest: None
pip: 18.0
setuptools: 39.0.1
Cython: None
numpy: 1.15.0
scipy: 1.1.0
pyarrow: None
xarray: None
IPython: 6.5.0
sphinx: None
patsy: None
dateutil: 2.7.3
pytz: 2018.5
blosc: None
bottleneck: None
tables: None
numexpr: None
feather: None
matplotlib: 2.2.3
openpyxl: None
xlrd: None
xlwt: None
xlsxwriter: None
lxml: None
bs4: None
html5lib: 1.0.1
sqlalchemy: None
pymysql: None
psycopg2: None
jinja2: 2.10
s3fs: None
fastparquet: None
pandas_gbq: None
pandas_datareader: None

[WillAyd]

Hmm this does look strange. If called with a dict this doesn't seem to exhibit the same behavior:

In [14]: df.replace(1, 10).dtypes                                               
Out[14]: 
first     float64
second     object
third      object
dtype: object

In [15]: df.replace({1: 10}).dtypes                                             
Out[15]: 
first      float64
second    category
third     category
dtype: object

Investigation and PRs certainly welcome

import pandas as pd 

>>> pd.__version__
'0.23.4'
>>> df = pd.DataFrame(
    [[1., 10, True, 'x'],
    [2, 20, False, 'y']],
    columns=['float', 'int', 'bool', 'str'])
>>> df.dtypes
float    float64 
int        int64 
bool        bool 
str       object 
dtype: object    
>>> df2 = df.astype({
    'float': 'category',
    'int': 'category',
    'bool': 'category'
    'str': 'category'
    }) 
>>> df2.dtypes # all columns' type is category
float    category
int      category
bool     category
str      category
dtype: object
>>> df2_replace_float_column = df2.replace(1, 10)
>>> df2_replace_float_column.dtypes # just int no change...
float     float64
int      category
bool       object
str        object
dtype: object
>>> df2_replace_float_column_new = df2.replace({1:10})
>>> df2_replace_float_column_new.dtypes # bool changed
float       int64   
int      category   
bool        int64   
str      category   
dtype: object       

So using dict may not be the best way.
A simple way is :

>> df2_new = df.replace({1:10}).astype({
    'float': 'category',
    'int': 'category',
    'bool': 'category'
    'str': 'category'
    }) 
>>> df2_new.dtypes
float    category
int      category
bool     category
str      category
dtype: object

replace() created a new DataFrame object:

>>>  id(df2) == id(df2.replace({1:10}))
 False

[fuglede]

When something happens here, you may want to add the details to this related StackOverflow question.

[mroeschke]

Looks like the categorical dtypes are kept now. Could use a test

In [103]: >>> df = pd.DataFrame(
     ...: ...     [[1., 'A', 'x'], [2, 'B', 'y'], [3, 'C', 'z']],
     ...: ...     columns=['first', 'second', 'third']
     ...: ... ).astype({'second': 'category', 'third': 'category'})
     ...:
     ...: # Replace int values
     ...: >>> df = df.replace(1, 10)
     ...:
     ...: # Both categorical columns turned into object...
     ...: >>> df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 3 columns):
 #   Column  Non-Null Count  Dtype
---  ------  --------------  -----
 0   first   3 non-null      float64
 1   second  3 non-null      category
 2   third   3 non-null      category
dtypes: category(2), float64(1)
memory usage: 366.0 bytes

In [104]: >>> df = pd.DataFrame(
     ...: ...     [['A', 'x'], ['B', 'y'], ['C', 'z']],
     ...: ...     columns=['first', 'second',]
     ...: ... ).astype({'first': 'category', 'second': 'category'})
     ...:
     ...: >>> df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 2 columns):
 #   Column  Non-Null Count  Dtype
---  ------  --------------  -----
 0   first   3 non-null      category
 1   second  3 non-null      category
dtypes: category(2)
memory usage: 342.0 bytes

In [105]: >>> df = df.replace('A', 'B')
     ...:
     ...: # Dtype of column `second` becomes object
     ...: >>> df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 3 entries, 0 to 2
Data columns (total 2 columns):
 #   Column  Non-Null Count  Dtype
---  ------  --------------  -----
 0   first   3 non-null      category
 1   second  3 non-null      category
dtypes: category(2)
memory usage: 334.0 bytes

In [106]: pd.__version__
Out[106]: '1.1.0.dev0+1974.g0159cba6e'