How to swap a group of column headings with their values in Pandas

Question

I have the following data frame:

a1  | a2  | a3  | a4 
--------------------- 
Bob | Cat | Dov | Edd 
Cat | Dov | Bob | Edd
Edd | Cat | Dov | Bob

and I want to convert it to

Bob | Cat | Dov | Edd
---------------------
a1  | a2  | a3  | a4
a3  | a1  | a2  | a4
a4  | a2  | a3  | a1

Note that the number of columns equals the number of unique values, and the number and order of rows are preserved

Answer 1

1) Required approach:

A faster implementation would be to sort the values of the dataframe and align the columns accordingly based on it's obtained indices after np.argsort.

pd.DataFrame(df.columns[np.argsort(df.values)], df.index, np.unique(df.values))

Applying np.argsort gives us the data we are looking for:

df.columns[np.argsort(df.values)]
Out[156]:
Index([['a1', 'a2', 'a3', 'a4'], ['a3', 'a1', 'a2', 'a4'],
       ['a4', 'a2', 'a3', 'a1']],
      dtype='object')

---

2) Slow generalized approach:

More generalized approach while at the cost of some speed / efficiency would be to use apply after creating a dict mapping of the strings/values present in the dataframe with their corresponding column names.

Use a dataframe constructor later after converting the obtained series to their list representation.

pd.DataFrame(df.apply(lambda s: dict(zip(pd.Series(s), pd.Series(s).index)), 1).tolist()) 

---

3) Faster generalized approach:

After obtaining a list of dictionaries from df.to_dict + orient='records', we need to swap it's respective key and value pairs while iterating through them in a loop.

pd.DataFrame([{val:key for key, val in d.items()} for d in df.to_dict('r')])

---

Sample test case:

df = df.assign(a5=['Foo', 'Bar', 'Baz'])

Both these approaches produce:

---

@piRSquared EDIT ¹

generalized solution

def nic(df):
    v = df.values
    n, m = v.shape
    u, inv = np.unique(v, return_inverse=1)
    i = df.index.values
    c = df.columns.values
    r = np.empty((n, len(u)), dtype=c.dtype)
    r[i.repeat(m), inv] = np.tile(c, n)
    return pd.DataFrame(r, i, u)

¹ _{I would like to thank user @piRSquared for coming up with a really fast and generalized numpy based alternative soln.}

Answer 2

numpy + pandas

v = df.values
n, m = v.shape
i = df.index.values
c = df.columns.values

# create series with values that were column values
# create multi index with first level from existing index
# and second level from flattened existing values
# then unstack
pd.Series(
    np.tile(c, n),
    [i.repeat(m), v.ravel()]
).unstack()

  Bob Cat Dov Edd
0  a1  a2  a3  a4
1  a3  a1  a2  a4
2  a4  a2  a3  a1

Answer 3

You can reshape it with stack and unstack with a swapping of the values and index:

df_swap = (df.stack()                     # reshape the data frame to long format
             .reset_index(level = 1)      # set the index(column headers) as a new column
             .set_index(0, append=True)   # set the values as index
             .unstack(level=1))           # reshape the data frame to wide format

df_swap.columns = df_swap.columns.get_level_values(1)   # drop level 0 in the column index
df_swap