考虑以下代码

import pandas as pd
df = pd.DataFrame({'col_1' : [1, 2, 3, 4],\
                   'col_2' : ['a', 'b', 'c', 'd'],\
                   'col_3' : ['hey', 'ho', 'banana', 'go']})

col = df['col_1'].astype(str) + '_' + \
      df['col_2'].astype(str) + '_' + \
      df['col_3'].astype(str)

col
Out[12]:
0       1_a_hey
1        2_b_ho
2    3_c_banana
3        4_d_go
dtype: object


有人能想到使用数组col作为输入来生成col_names = ['col_1', 'col_2', 'col_3']的oneliner吗?

那是col_sum = something_smart(col_names)

显然,例如different_col_set = ['col_2', 'col_3']

something_smart(different_col_set)
Out[13]:
0         a_hey
1          b_ho
2      c_banana
3          d_go
dtype: object


关键是col_names是一个数组,其中包含数据框的列名称的任何子集。

最佳答案

选项1]使用apply可以'_'.join

In [5521]: df[col_names].astype(str).apply('_'.join, axis=1)
Out[5521]:
0       1_a_hey
1        2_b_ho
2    3_c_banana
3        4_d_go
dtype: object


和,

In [5523]: df[different_col_set].astype(str).apply('_'.join, axis=1)
Out[5523]:
0       a_hey
1        b_ho
2    c_banana
3        d_go
dtype: object


选项2]使用reduce的速度比在这种情况下应用的速度更快。

In [5527]: reduce(lambda x, y: x + '_' + y, [df[c].astype(str) for c in col_names])
Out[5527]:
0       1_a_hey
1        2_b_ho
2    3_c_banana
3        4_d_go
dtype: object

In [5528]: reduce(lambda x, y: x + '_' + y, [df[c].astype(str) for c in different_col_set])
Out[5528]:
0       a_hey
1        b_ho
2    c_banana
3        d_go
dtype: object


类似于reduce(lambda x, y: x.astype(str) + '_' +y.astype(str), [df[x] for x in col_names])



时机

In [5556]: df.shape
Out[5556]: (10000, 3)

In [5553]: %timeit reduce(lambda x, y: x + '_' + y, [df[c].astype(str) for c in col_names])
10 loops, best of 3: 21.7 ms per loop

In [5554]: %timeit reduce(lambda x, y: x.astype(str) + '_' +y.astype(str), [df[x] for x in col_names])
10 loops, best of 3: 22.3 ms per loop

In [5555]: %timeit df[col_names].astype(str).apply('_'.join, axis=1)
1 loop, best of 3: 254 ms per loop

10-07 20:19