Rename MultiIndex columns in Pandas

In pandas 0.21.0+ use parameter level=1:

d = dict(zip(df.columns.levels[1], ["b1", "c1", "f1"]))
print (d)
{'c': 'c1', 'b': 'b1', 'f': 'f1'}

df = df.rename(columns=d, level=1)
print (df)
     a         d
    b1   c1   f1
0    1    2    3
1   10   20   30
2  100  200  300

You can use pandas.DataFrame.rename() directly

Say you have the following dataframe

print(df)

     a         d
     b    c    f
0    1    2    3
1   10   20   30
2  100  200  300

df = df.rename(columns={'f': 'f1', 'd': 'd1'})

print(df)

     a        d1
     b    c   f1
0    1    2    3
1   10   20   30
2  100  200  300

You see, column name mapper doesn't relate with level.

Say you have the following dataframe

     a         d
     b    f    f
0    1    2    3
1   10   20   30
2  100  200  300

If you want to rename the f under a, you can do

df.columns = df.columns.values
df.columns = pd.MultiIndex.from_tuples(df.rename(columns={('a', 'f'): ('a', 'af')}))
# or in one line
df.columns = pd.MultiIndex.from_tuples(df.set_axis(df.columns.values, axis=1)
                                       .rename(columns={('a', 'f'): ('a', 'af')}))

print(df)

     a         d
     b   af    f
0    1    2    3
1   10   20   30
2  100  200  300

There is also index.set_names (code)

df.index.set_names(["b1", "c1", "f1"], inplace=True)

Another thing you can't do is df.rename(columns={('d', 'f'): ('e', 'g')}), even though it seems correct. In other words: .rename() does not do what one expects, <...>

-- Lukas at comment

The "hacky" way is something like this (as far as for pandas 1.0.5)

def rename_columns(df, columns, inplace=False):
    """Rename dataframe columns.

    Parameters
    ----------
    df : pandas.DataFrame
        Dataframe.
    columns : dict-like
        Alternative to specifying axis. If `df.columns` is
        :obj: `pandas.MultiIndex`-object and has a few levels, pass equal-size tuples.

    Returns
    -------
    pandas.DataFrame or None
        Returns dataframe with modifed columns or ``None`` (depends on `inplace` parameter value).
    
    Examples
    --------
    >>> columns = pd.Index([1, 2, 3])
    >>> df = pd.DataFrame([[1, 2, 3], [10, 20, 30]], columns=columns)
    ...     1   2   3
    ... 0   1   2   3
    ... 1  10  20  30
    >>> rename_columns(df, columns={1 : 10})
    ...    10   2   3
    ... 0   1   2   3
    ... 1  10  20  30
    
    MultiIndex
    
    >>> columns = pd.MultiIndex.from_tuples([("A0", "B0", "C0"), ("A1", "B1", "C1"), ("A2", "B2", "")])
    >>> df = pd.DataFrame([[1, 2, 3], [10, 20, 30]], columns=columns)
    >>> df
    ...    A0  A1  A2
    ...    B0  B1  B2
    ...    C0  C1
    ... 0   1   2   3
    ... 1  10  20  30
    >>> rename_columns(df, columns={("A2", "B2", "") : ("A3", "B3", "")})
    ...    A0  A1  A3
    ...    B0  B1  B3
    ...    C0  C1
    ... 0   1   2   3
    ... 1  10  20  30
    """
    columns_new = []
    for col in df.columns.values:
        if col in columns:
            columns_new.append(columns[col])
        else:
            columns_new.append(col)
    columns_new = pd.Index(columns_new, tupleize_cols=True)

    if inplace:
        df.columns = columns_new
    else:
        df_new = df.copy()
        df_new.columns = columns_new
        return df_new

So just

>>> df = pd.DataFrame([[1,2,3], [10,20,30], [100,200,300]])
>>> df.columns = pd.MultiIndex.from_tuples((("a", "b"), ("a", "c"), ("d", "f")))
>>> rename_columns(df, columns={('d', 'f'): ('e', 'g')})
...      a         e
...      b    c    g
... 0    1    2    3
... 1   10   20   30
... 2  100  200  300

What does the pandas-team think about this? Why is this behavior not default?

Another way to do that is with pandas.Series.map and a lambda function as follows

df.columns = df.columns.map(lambda x: (x[0], "e") if x[1] == "f" else x)

[Out]:
     a         d
     b    c    e
0    1    2    3
1   10   20   30
2  100  200  300

Using dicts to rename tuples

Since multi-index stores values as tuples, and python dicts accept tuples as keys and values, we can replace them using a dict.

mapping_dict = {("d","f"):("d","e")}

# Dictionary allows using tuples as keys and values
def rename_tuple(tuple_, dict_):
    """Replaces tuple if present in tuple dict"""
    if tuple_ in dict_.keys():
        return dict_[tuple_]
    return tuple_

# Rename chosen elements from list of tuples from df.columns
altered_index_list = [rename_tuple(tuple_,mapping_dict) for tuple_ in df.columns.to_list()]

# Update columns with new renamed columns
df.columns = pd.Index(altered_index_list)

Which returns the intended df

     a         d
     b    c    e
0    1    2    3
1   10   20   30
2  100  200  300

Aggregating in a function

This could then be aggregated in a function to simplify things

def rename_multi_index(index,mapper):
    """Renames pandas multi_index"""
    return pd.Index([rename_tuple(tuple_,mapper) for tuple_ in index])

# And now simply do
df.columns = rename_multi_index(df.columns,mapping_dict)

Another mapping option renaming "f" to "e" (the original question).

Starting with the example dataframe:

# First define a mapping (this could hold more key:value pairs)
mapper = {("d", "f"): ("d", "e")}

# Then loop over all columns and replace whenever a match is found
df.columns = pd.MultiIndex.from_tuples([mapper[x] if x in mapper else x for x in df.columns])

The result is then: