访问pandas.DataFrame列的不同方法

Question

What is the difference between df[columns] and df.loc[:,columns] , both as lvalue and rvalue?

They seem to be interchangeable from the behavioral POV:

>>> df = pd.DataFrame({'x':[1,2,3],'y':['a','b','c']})
>>> df[['x']].equals(df.loc[:,['x']])
True
>>> df.loc[:,'z'] = df.x.apply(str) + df.y
>>> df['a'] = df.x.apply(str) + df.y
>>> df
   x  y   z   a
0  1  a  1a  1a
1  2  b  2b  2b
2  3  c  3c  3c

I know there is a document somewhere answering this in excruciating details (and I am sure I even saw it once, but a link would be nice), but I am looking at an "executive summary", so to speak.

Specifically : is one a shortcut for the other, or there is some semantic difference?

PS. This is prompted by the message

~/.virtualenvs/wilbur/lib/python2.7/site-packages/pandas/core/indexing.py:465: SettingWithCopyWarning: A value is trying to be set on a copy of a slice from a DataFrame. Try using .loc[row_indexer,col_indexer] = value instead

Answer 1

what about timing comparison for a 300K rows DF?

In [22]: big = pd.concat([df] * 10**5, ignore_index=True)

In [23]: %timeit -n 1 -r 1 big['n1'] = big.x.apply(str) + big.y
1 loop, best of 1: 266 ms per loop

In [24]: %timeit -n1 -r 1 big.ix[:, 'n2'] = big.x.apply(str) + big.y
1 loop, best of 1: 317 ms per loop

In [25]: %timeit -n 1 -r 1 big.loc[:, 'n3'] = big.x.apply(str) + big.y
1 loop, best of 1: 333 ms per loop

In [26]: %timeit -n 1 -r 1 big.insert(len(big.columns), 'n4', big.x.apply(str) + big.y)
1 loop, best of 1: 266 ms per loop

In [27]: big.shape
Out[27]: (300000, 6)

In [28]: big.head()
Out[28]:
   x  y  n1  n2  n3  n4
0  1  a  1a  1a  1a  1a
1  2  b  2b  2b  2b  2b
2  3  c  3c  3c  3c  3c
3  1  a  1a  1a  1a  1a
4  2  b  2b  2b  2b  2b