公认的解决方案对于大量数据来说将会非常慢。获得最多赞数的解决方案读起来有点困难，而且处理数字数据也很慢。如果每个新列都可以独立于其他列计算，那么我将直接分配它们，而不使用apply。

假字符数据的例子

在DataFrame中创建100,000个字符串

df = pd.DataFrame(np.random.choice(['he jumped', 'she ran', 'they hiked'],
                                   size=100000, replace=True),
                  columns=['words'])
df.head()
        words
0     she ran
1     she ran
2  they hiked
3  they hiked
4  they hiked

假设我们想提取一些文本特征，就像在最初的问题中所做的那样。例如，让我们提取第一个字符，计算字母“e”的出现次数，并将短语大写。

df['first'] = df['words'].str[0]
df['count_e'] = df['words'].str.count('e')
df['cap'] = df['words'].str.capitalize()
df.head()
        words first  count_e         cap
0     she ran     s        1     She ran
1     she ran     s        1     She ran
2  they hiked     t        2  They hiked
3  they hiked     t        2  They hiked
4  they hiked     t        2  They hiked

计时

%%timeit
df['first'] = df['words'].str[0]
df['count_e'] = df['words'].str.count('e')
df['cap'] = df['words'].str.capitalize()
127 ms ± 585 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

def extract_text_features(x):
    return x[0], x.count('e'), x.capitalize()

%timeit df['first'], df['count_e'], df['cap'] = zip(*df['words'].apply(extract_text_features))
101 ms ± 2.96 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

令人惊讶的是，通过遍历每个值可以获得更好的性能

%%timeit
a,b,c = [], [], []
for s in df['words']:
    a.append(s[0]), b.append(s.count('e')), c.append(s.capitalize())

df['first'] = a
df['count_e'] = b
df['cap'] = c
79.1 ms ± 294 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

另一个假数字数据的例子

创建100万个随机数并从上面测试幂函数。

df = pd.DataFrame(np.random.rand(1000000), columns=['num'])


def powers(x):
    return x, x**2, x**3, x**4, x**5, x**6

%%timeit
df['p1'], df['p2'], df['p3'], df['p4'], df['p5'], df['p6'] = \
       zip(*df['num'].map(powers))
1.35 s ± 83.6 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

为每一列赋值速度快25倍，可读性强:

%%timeit 
df['p1'] = df['num'] ** 1
df['p2'] = df['num'] ** 2
df['p3'] = df['num'] ** 3
df['p4'] = df['num'] ** 4
df['p5'] = df['num'] ** 5
df['p6'] = df['num'] ** 6
51.6 ms ± 1.9 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

我在这里也做了类似的回答，并详细说明了为什么申请通常不是正确的选择。

基于user1827356的答案，你可以使用df.merge一次性完成赋值:

df.merge(df.textcol.apply(lambda s: pd.Series({'feature1':s+1, 'feature2':s-1})), 
    left_index=True, right_index=True)

    textcol  feature1  feature2
0  0.772692  1.772692 -0.227308
1  0.857210  1.857210 -0.142790
2  0.065639  1.065639 -0.934361
3  0.819160  1.819160 -0.180840
4  0.088212  1.088212 -0.911788

编辑: 请注意内存消耗大，速度慢:https://ys-l.github.io/posts/2015/08/28/how-not-to-use-pandas-apply/ !

对于95%的用例来说，这是正确且最简单的方法:

>>> df = pd.DataFrame(zip(*[range(10)]), columns=['num'])
>>> df
    num
0    0
1    1
2    2
3    3
4    4
5    5

>>> def example(x):
...     x['p1'] = x['num']**2
...     x['p2'] = x['num']**3
...     x['p3'] = x['num']**4
...     return x

>>> df = df.apply(example, axis=1)
>>> df
    num  p1  p2  p3
0    0   0   0    0
1    1   1   1    1
2    2   4   8   16
3    3   9  27   81
4    4  16  64  256

在2020年，我使用apply()参数result_type='expand'

applied_df = df.apply(lambda row: fn(row.text), axis='columns', result_type='expand')
df = pd.concat([df, applied_df], axis='columns')

我有一个更复杂的情况，数据集有一个嵌套结构:

import json
data = '{"TextID":{"0":"0038f0569e","1":"003eb6998d","2":"006da49ea0"},"Summary":{"0":{"Crisis_Level":["c"],"Type":["d"],"Special_Date":["a"]},"1":{"Crisis_Level":["d"],"Type":["a","d"],"Special_Date":["a"]},"2":{"Crisis_Level":["d"],"Type":["a"],"Special_Date":["a"]}}}'
df = pd.DataFrame.from_dict(json.loads(data))
print(df)

输出:

        TextID                                            Summary
0  0038f0569e  {'Crisis_Level': ['c'], 'Type': ['d'], 'Specia...
1  003eb6998d  {'Crisis_Level': ['d'], 'Type': ['a', 'd'], 'S...
2  006da49ea0  {'Crisis_Level': ['d'], 'Type': ['a'], 'Specia...

Summary列包含dict对象，所以我使用apply和from_dict和stack来提取每一行的dict:

df2 = df.apply(
    lambda x: pd.DataFrame.from_dict(x[1], orient='index').stack(), axis=1)
print(df2)

输出:

    Crisis_Level Special_Date Type     
                0            0    0    1
0            c            a    d  NaN
1            d            a    a    d
2            d            a    a  NaN

看起来不错，但缺少TextID列。为了得到TextID列回来，我尝试了三种方法:

Modify apply to return multiple columns: df_tmp = df.copy() df_tmp[['TextID', 'Summary']] = df.apply( lambda x: pd.Series([x[0], pd.DataFrame.from_dict(x[1], orient='index').stack()]), axis=1) print(df_tmp) output: TextID Summary 0 0038f0569e Crisis_Level 0 c Type 0 d Spec... 1 003eb6998d Crisis_Level 0 d Type 0 a ... 2 006da49ea0 Crisis_Level 0 d Type 0 a Spec... But this is not what I want, the Summary structure are flatten. Use pd.concat: df_tmp2 = pd.concat([df['TextID'], df2], axis=1) print(df_tmp2) output: TextID (Crisis_Level, 0) (Special_Date, 0) (Type, 0) (Type, 1) 0 0038f0569e c a d NaN 1 003eb6998d d a a d 2 006da49ea0 d a a NaN Looks fine, the MultiIndex column structure are preserved as tuple. But check columns type: df_tmp2.columns output: Index(['TextID', ('Crisis_Level', 0), ('Special_Date', 0), ('Type', 0), ('Type', 1)], dtype='object') Just as a regular Index class, not MultiIndex class. use set_index: Turn all columns you want to preserve into row index, after some complicated apply function and then reset_index to get columns back: df_tmp3 = df.set_index('TextID') df_tmp3 = df_tmp3.apply( lambda x: pd.DataFrame.from_dict(x[0], orient='index').stack(), axis=1) df_tmp3 = df_tmp3.reset_index(level=0) print(df_tmp3) output: TextID Crisis_Level Special_Date Type 0 0 0 1 0 0038f0569e c a d NaN 1 003eb6998d d a a d 2 006da49ea0 d a a NaN Check the type of columns df_tmp3.columns output: MultiIndex(levels=[['Crisis_Level', 'Special_Date', 'Type', 'TextID'], [0, 1, '']], codes=[[3, 0, 1, 2, 2], [2, 0, 0, 0, 1]])

因此，如果apply函数将返回MultiIndex列，并且希望保留它，则可能需要尝试第三种方法。

在另外两个类似的问题上也有相同的答案。我更喜欢这样做的方式是将函数的返回值打包成一个系列:

def f(x):
    return pd.Series([x**2, x**3])

然后使用apply创建单独的列，如下所示:

df[['x**2','x**3']] = df.apply(lambda row: f(row['x']), axis=1)