I've noticed three methods of selecting a column in a Pandas DataFrame:
First method of selecting a column using loc:
df_new = df.loc[:, 'col1']
Second method - seems simpler and faster:
df_new = df['col1']
Third method - most convenient:
df_new = df.col1
Is there a difference between these three methods? I don't think so, in which case I'd rather use the third method.
I'm mostly curious as to why there appear to be three methods for doing the same thing.
In the following situations, they behave the same:
df['A'] is the same as df.loc[:, 'A'] -> selects column A)df[['A', 'B', 'C']] is the same as df.loc[:, ['A', 'B', 'C']] -> selects columns A, B and C)df[1:3] is the same as df.iloc[1:3] -> selects rows 1 and 2. Note, however, if you slice rows with loc, instead of iloc, you'll get rows 1, 2 and 3 assuming you have a RangeIndex. See details here.)However, [] does not work in the following situations:
df.loc[row_label]df.loc[[row_label1, row_label2]]df.loc[:, 'A':'C']These three cannot be done with [].
More importantly, if your selection involves both rows and columns, then assignment becomes problematic.
df[1:3]['A'] = 5
This selects rows 1 and 2 then selects column 'A' of the returning object and assigns value 5 to it. The problem is, the returning object might be a copy so this may not change the actual DataFrame. This raises SettingWithCopyWarning. The correct way of making this assignment is:
df.loc[1:3, 'A'] = 5
With .loc, you are guaranteed to modify the original DataFrame. It also allows you to slice columns (df.loc[:, 'C':'F']), select a single row (df.loc[5]), and select a list of rows (df.loc[[1, 2, 5]]).
Also note that these two were not included in the API at the same time. .loc was added much later as a more powerful and explicit indexer. See unutbu's answer for more detail.
Note: Getting columns with [] vs . is a completely different topic. . is only there for convenience. It only allows accessing columns whose names are valid Python identifiers (i.e. they cannot contain spaces, they cannot be composed of numbers...). It cannot be used when the names conflict with Series/DataFrame methods. It also cannot be used for non-existing columns (i.e. the assignment df.a = 1 won't work if there is no column a). Other than that, . and [] are the same.
df[1:3]['A'] = 5 to be a copy or not?df.__getitem__(...), and under the hood, the memory layout of the stored array can result in a view or a copy. In general, when you work on a dataframe with a single dtype, you get a view but that's not guaranteed. I believe they are working on a new approach instead of using BlockManager which is the main source of these issues.
loc is specially useful when the index is not numeric (e.g. a DatetimeIndex) because you can get rows with particular labels from the index:
df.loc['2010-05-04 07:00:00']
df.loc['2010-1-1 0:00:00':'2010-12-31 23:59:59 ','Price']
However [] is intended to get columns with particular names:
df['Price']
With [] you can also filter rows, but it is more elaborated:
df[df['Date'] < datetime.datetime(2010,1,1,7,0,0)]['Price']
df_augmented['date_of_birth'] = pd.to_datetime(df_augmented.date_of_birth, format='mixed') working, but not df_augmented.loc[:, 'date_of_birth'] = pd.to_datetime(df_augmented.date_of_birth, format='mixed') - the latter gives an Object column, the former a datetime64[ns]; why? are there differences between df['col'] and df[:, 'col'] in this case?If you're confused which of these approaches is (at least) the recommended one for your use-case, take a look at this brief instructions from pandas tutorial:
When selecting subsets of data, square brackets [] are used.
Inside these brackets, you can use a single column/row label, a list of column/row labels, a slice of labels, a conditional expression or a colon.
Select specific rows and/or columns using loc when using the row and
column names
Select specific rows and/or columns using iloc when using the
positions in the table
You can assign new values to a selection based on loc/iloc.
I highlighted some of the points to make their use-case differences even more clear.
There seems to be a difference between df.loc[] and df[] when you create dataframe with multiple columns.
You can refer to this question: Is there a nice way to generate multiple columns using .loc?
Here, you can't generate multiple columns using df.loc[:,['name1','name2']] but you can do by just using double bracket df[['name1','name2']]. (I wonder why they behave differently.)
There is an important difference in assignment. These two assignments are not equivalent:
df['j'] = pd.Series(...)
df.loc[:,'j'] = pd.Series(...)
Indexing with [...] replaces the series. The new series comes with its own dtype, the existing dtype is discarded.
Indexing with loc[...] replaces values, but reuses the existing series and its dtype, upcasting may occur to fit new values.
See how the old int32 is ignored when using [...]:
import pandas as pd
import numpy.random as npr
n = 4
# NumPy returns dtype int32
df = pd.DataFrame({
'j': npr.randint(1, 10, n),
'k': npr.randint(1, 10, n)})
print(df.dtypes)
# uint8 series
s = pd.Series(npr.randint(1, 10, n), dtype='uint8')
# Using [...]: uint8 series replaces uint32 series
df2 = df.copy()
df2['j'] = s
print(df2.dtypes)
# Using loc[...]: uint8 data upcasted to existing uint32
df3 = df.copy()
df3.loc[:,'j'] = s
print(df3.dtypes)
j int32 ⇠ original dtype
k int32
dtype: object
j uint8 ⇠ with [...]
k int32
dtype: object
j int32 ⇠ with loc[...]
k int32
dtype: object
df.col1? All three of these are essentially equivalent for the very simple case of selecting a column..locwill let you do much more than select a column. Possible duplicate of stackoverflow.com/questions/31593201/…df.sum, what happens? (spoiler alert, nothing useful, althoughdf.sum()still works luckily) So 3rd way should be seen as a shortcut that is fine, but need to be careful with