Collapsing rows in a Pandas dataframe if all rows

问题:

I have following DF

         col1  |  col2   | col3   | col4   | col5  | col6
    0    -     |   15.0  |  -     |  -     |   -   |  -
    1    -     |   -     |  -     |  -     |   -   |  US
    2    -     |   -     |  -     |  Large |   -   |  -
    3    ABC1  |   -     |  -     |  -     |   -   |  -
    4    -     |   -     |  24RA  |  -     |   -   |  -
    5    -     |   -     |  -     |  -     |   345 |  -

I want to collapse rows into one as follows

    output DF:
         col1  |  col2    | col3   | col4   | col5  | col6
    0    ABC1  |   15.0   |  24RA  |  Large |   345 |  US

I do not want to iterate over columns but want to use pandas to achieve this.

回答1:

Option 0
Super Simple

pd.concat([pd.Series(df[c].dropna().values, name=c) for c in df], axis=1)

   col1  col2  col3   col4   col5 col6
0  ABC1  15.0  24RA  Large  345.0   US

---

Can we handle more than one value per column?
Sure we can!

df.loc[2, 'col3'] = 'Test'

   col1  col2  col3   col4   col5 col6
0  ABC1  15.0  Test  Large  345.0   US
1   NaN   NaN  24RA    NaN    NaN  NaN

Option 1
Generalized solution using np.where like a surgeon

v = df.values
i, j = np.where(np.isnan(v))

s = pd.Series(v[i, j], df.columns[j])

c = s.groupby(level=0).cumcount()
s.index = [c, s.index]
s.unstack(fill_value='-')  # <-- don't fill to get NaN

   col1  col2  col3   col4 col5 col6
0  ABC1  15.0  24RA  Large  345   US

---

df.loc[2, 'col3'] = 'Test'

v = df.values
i, j = np.where(np.isnan(v))

s = pd.Series(v[i, j], df.columns[j])

c = s.groupby(level=0).cumcount()
s.index = [c, s.index]
s.unstack(fill_value='-')  # <-- don't fill to get NaN

   col1  col2  col3   col4 col5 col6
0  ABC1  15.0  Test  Large  345   US
1     -     -  24RA      -    -    -

---

Option 2
mask to make nulls then stack to get rid of them

Or we could have

# This should work even if `'-'` are NaN
# but you can skip the `.mask(df == '-')`
s = df.mask(df == '-').stack().reset_index(0, drop=True)
c = s.groupby(level=0).cumcount()
s.index = [c, s.index]
s.unstack(fill_value='-')

   col1  col2  col3   col4 col5 col6
0  ABC1  15.0  Test  Large  345   US
1     -     -  24RA      -    -    -

回答2:

You can use max, but you need to convert the null values in the string-valued columsn (which is a bit ugly unfortunately)

>>> df = pd.DataFrame({'col1':[np.nan, "ABC1"], 'col2':[15.0, np.nan]})

>>> df.apply(lambda c: c.fillna('') if c.dtype is np.dtype('O') else c).max()
col1    ABC1
col2      15
dtype: object

You could also you a combination of backfill and forwardfill to fill in the gaps, this could be useful if only want to apply this to some of your columns:

>>> df.apply(lambda c: c.fillna(method='bfill').fillna(method='ffill'))