堆叠时设置列名 pandas DataFrame

Question

When stacking a pandas DataFrame , a Series is returned. Normally after I stack a DataFrame , I convert it back into a DataFrame . However, the default names coming from the stacked data make renaming the columns a bit hacky. What I'm looking for is an easier/built-in way to give columns sensible names after stacking.

Eg, for the following DataFrame :

In [64]: df = pd.DataFrame({'id':[1,2,3], 
    ...:                    'date':['2015-09-31']*3, 
    ...:                    'value':[100, 95, 42], 
    ...:                    'value2':[200, 57, 27]}).set_index(['id','date'])

In [65]: df
Out[65]: 
               value  value2
id date                     
1  2015-09-31    100     200
2  2015-09-31     95      57
3  2015-09-31     42      27

I stack and convert it back to a DataFrame like so:

In [68]: df.stack().reset_index()
Out[68]: 
   id        date level_2    0
0   1  2015-09-31   value  100
1   1  2015-09-31  value2  200
2   2  2015-09-31   value   95
3   2  2015-09-31  value2   57
4   3  2015-09-31   value   42
5   3  2015-09-31  value2   27

So in order to name these columns appropriately I would need to do something like this:

In [72]: stacked = df.stack()

In [73]: stacked
Out[73]: 
id  date              
1   2015-09-31  value     100
                value2    200
2   2015-09-31  value      95
                value2     57
3   2015-09-31  value      42
                value2     27
dtype: int64

In [74]: stacked.index.set_names('var_name', level=len(stacked.index.names)-1, inplace=True)

In [88]: stacked.reset_index().rename(columns={0:'value'})
Out[88]: 
   id        date var_name  value
0   1  2015-09-31    value    100
1   1  2015-09-31   value2    200
2   2  2015-09-31    value     95
3   2  2015-09-31   value2     57
4   3  2015-09-31    value     42
5   3  2015-09-31   value2     27

Ideally, the solution would look something like this:

df.stack(new_index_name='var_name', new_col_name='value')

But looking at the docs it doesn't look like stack takes any such arguments. Is there an easier/built-in way in pandas to deal with this workflow?

Answer 1

So here's one way that you may find a bit cleaner, using the fact that columns and Series can also carry names.

In [45]: df
Out[45]: 
               value  value2
id date                     
1  2015-09-31    100     200
2  2015-09-31     95      57
3  2015-09-31     42      27

In [46]: df.columns.name = 'var_name'

In [47]: s = df.stack()

In [48]: s.name = 'value'

In [49]: s.reset_index()
Out[49]: 
   id        date var_name  value
0   1  2015-09-31    value    100
1   1  2015-09-31   value2    200
2   2  2015-09-31    value     95
3   2  2015-09-31   value2     57
4   3  2015-09-31    value     42
5   3  2015-09-31   value2     27

Answer 2

pd.melt is often useful for converting DataFrames from "wide" to "long" format. You could use pd.melt here if you convert the id and date index levels to columns first:

In [56]: pd.melt(df.reset_index(), id_vars=['id', 'date'], value_vars=['value', 'value2'], var_name='var_name', value_name='value')
Out[56]: 
   id        date var_name  value
0   1  2015-09-31    value    100
1   2  2015-09-31    value     95
2   3  2015-09-31    value     42
3   1  2015-09-31   value2    200
4   2  2015-09-31   value2     57
5   3  2015-09-31   value2     27

Answer 3

A pipe-ing friendly alternative to chrisb's answer:

df.stack().rename_axis(['id', 'date', 'var_name']).rename('value').reset_index()

And if explicit is better than implicit:

(
    df
    .stack()
    .rename_axis(index={'id': 'id', 'date': 'date', None: 'var_name'})
    .rename('value')
    .reset_index()
)

When using the dict mapper, you can skip the names which should stay the same:

df.stack().rename_axis(index={None: 'var_name'}).rename('value').reset_index()

Answer 4

Why not something like this? Sometimes melt is great, but sometimes you want to keep your index, and/or you want to have an index on that new column. This is like @krassowski's answer, but it doesn't require you to know the names of df's indices in advance.

df.stack().rename_axis([*df.index.names, "var_name"]).rename("value")

Answer 5

To avoid a phantom column name when calling stack , just rename the column axis beforehand:

df = pd.DataFrame({col: range(3) for col in list("ABC")})
df.rename_axis(columns="lol_goodbye_columns").stack()

   lol_goodbye_columns
0  A                      0
   B                      0
   C                      0
1  A                      1
   B                      1
   C                      1
2  A                      2
   B                      2
   C                      2
dtype: int64