提问人:LondonRob 提问时间:11/14/2013 最后编辑:cottontailLondonRob 更新时间:11/7/2023 访问量:1189545
如何像在 SQL 中那样使用“in”和“not in”过滤 Pandas 数据帧
How to filter Pandas dataframe using 'in' and 'not in' like in SQL
问:
如何实现 SQL 和 的等价物?IN
NOT IN
我有一个包含所需值的列表。场景如下:
df = pd.DataFrame({'country': ['US', 'UK', 'Germany', 'China']})
countries_to_keep = ['UK', 'China']
# pseudo-code:
df[df['country'] not in countries_to_keep]
我目前的执行方法如下:
df = pd.DataFrame({'country': ['US', 'UK', 'Germany', 'China']})
df2 = pd.DataFrame({'country': ['UK', 'China'], 'matched': True})
# IN
df.merge(df2, how='inner', on='country')
# NOT IN
not_in = df.merge(df2, how='left', on='country')
not_in = not_in[pd.isnull(not_in['matched'])]
但这似乎是一个可怕的笨拙。任何人都可以改进它吗?
答:
你可以使用 pd.系列.isin
。
对于“IN”使用:something.isin(somewhere)
或者对于“不在”:~something.isin(somewhere)
举个工作的例子:
>>> df
country
0 US
1 UK
2 Germany
3 China
>>> countries_to_keep
['UK', 'China']
>>> df.country.isin(countries_to_keep)
0 False
1 True
2 False
3 True
Name: country, dtype: bool
>>> df[df.country.isin(countries_to_keep)]
country
1 UK
3 China
>>> df[~df.country.isin(countries_to_keep)]
country
0 US
2 Germany
评论
df = pd.Series({'countries':['US','UK','Germany','China']})
df
DataFrame
countries
df[~df.countries.isin(countries)]
DataFrame
Series
countries
我通常对这样的行进行泛型过滤:
criterion = lambda row: row['countries'] not in countries
not_in = df[df.apply(criterion, axis=1)]
评论
我想过滤掉具有 dfProfilesBusIds BUSINESS_ID BUSINESS_ID的 dfbc 行
dfbc = dfbc[~dfbc['BUSINESS_ID'].isin(dfProfilesBusIds['BUSINESS_ID'])]
使用 .query() 方法的替代解决方案:
In [5]: df.query("country in @countries_to_keep")
Out[5]:
countries
1 UK
3 China
In [6]: df.query("country not in @countries_to_keep")
Out[6]:
countries
0 US
2 Germany
评论
countries
countries_to_keep
countries_to_keep
countries_to_keep
df = pd.DataFrame({'countries':['US','UK','Germany','China']})
countries = ['UK','China']
实施方式:
df[df.countries.isin(countries)]
在其他国家/地区实施:
df[df.countries.isin([x for x in np.unique(df.countries) if x not in countries])]
如何为 pandas DataFrame 实现“in”和“not in”?
Pandas 提供两种方法:Series.isin 和 DataFrame.isin
,分别用于 Series 和 DataFrame。
基于一列筛选数据帧(也适用于系列)
最常见的方案是对特定列应用条件以筛选 DataFrame 中的行。isin
df = pd.DataFrame({'countries': ['US', 'UK', 'Germany', np.nan, 'China']})
df
countries
0 US
1 UK
2 Germany
3 China
c1 = ['UK', 'China'] # list
c2 = {'Germany'} # set
c3 = pd.Series(['China', 'US']) # Series
c4 = np.array(['US', 'UK']) # array
Series.isin
接受各种类型作为输入。以下是获得所需内容的所有有效方法:
df['countries'].isin(c1)
0 False
1 True
2 False
3 False
4 True
Name: countries, dtype: bool
# `in` operation
df[df['countries'].isin(c1)]
countries
1 UK
4 China
# `not in` operation
df[~df['countries'].isin(c1)]
countries
0 US
2 Germany
3 NaN
# Filter with `set` (tuples work too)
df[df['countries'].isin(c2)]
countries
2 Germany
# Filter with another Series
df[df['countries'].isin(c3)]
countries
0 US
4 China
# Filter with array
df[df['countries'].isin(c4)]
countries
0 US
1 UK
对许多列进行筛选
有时,您需要在多个列上应用带有某些搜索词的“in”成员资格检查,
df2 = pd.DataFrame({
'A': ['x', 'y', 'z', 'q'], 'B': ['w', 'a', np.nan, 'x'], 'C': np.arange(4)})
df2
A B C
0 x w 0
1 y a 1
2 z NaN 2
3 q x 3
c1 = ['x', 'w', 'p']
要将条件应用于“A”和“B”列,请使用:isin
DataFrame.isin
df2[['A', 'B']].isin(c1)
A B
0 True True
1 False False
2 False False
3 False True
由此,要保留至少有一列为 True
的行,我们可以沿第一个轴使用:any
df2[['A', 'B']].isin(c1).any(axis=1)
0 True
1 False
2 False
3 True
dtype: bool
df2[df2[['A', 'B']].isin(c1).any(axis=1)]
A B C
0 x w 0
3 q x 3
请注意,如果要搜索每一列,只需省略列选择步骤并执行
df2.isin(c1).any(axis=1)
同样,若要保留 ALL 列为 True
的行,请以与以前相同的方式使用。all
df2[df2[['A', 'B']].isin(c1).all(axis=1)]
A B C
0 x w 0
值得注意的提及: , , list comprehensions (string data)numpy.isin
query
除了上述方法外,您还可以使用 numpy 等效项:numpy.isin
。
# `in` operation
df[np.isin(df['countries'], c1)]
countries
1 UK
4 China
# `not in` operation
df[np.isin(df['countries'], c1, invert=True)]
countries
0 US
2 Germany
3 NaN
为什么值得考虑?由于开销较低,NumPy 函数通常比其 pandas 等效函数快一些。由于这是一个不依赖于索引对齐的元素操作,因此在极少数情况下,此方法不适合替代 pandas 的 .isin
Pandas 例程在处理字符串时通常是迭代的,因为字符串操作很难矢量化。有很多证据表明,这里的列表理解会更快。
我们现在求助于检查。in
c1_set = set(c1) # Using `in` with `sets` is a constant time operation...
# This doesn't matter for pandas because the implementation differs.
# `in` operation
df[[x in c1_set for x in df['countries']]]
countries
1 UK
4 China
# `not in` operation
df[[x not in c1_set for x in df['countries']]]
countries
0 US
2 Germany
3 NaN
但是,指定起来要笨拙得多,因此除非您知道自己在做什么,否则不要使用它。
最后,这个答案中也涵盖了哪些内容。numexpr FTW!DataFrame.query
评论
从答案中整理可能的解决方案:
对于 IN:df[df['A'].isin([3, 6])]
对于不在:
df[-df["A"].isin([3, 6])]
df[~df["A"].isin([3, 6])]
df[df["A"].isin([3, 6]) == False]
df[np.logical_not(df["A"].isin([3, 6]))]
评论
logical_not
~
如果您想保持列表的顺序,这是一个技巧:
df = pd.DataFrame({'country': ['US', 'UK', 'Germany', 'China']})
countries_to_keep = ['Germany', 'US']
ind=[df.index[df['country']==i].tolist() for i in countries_to_keep]
flat_ind=[item for sublist in ind for item in sublist]
df.reindex(flat_ind)
country
2 Germany
0 US
我的 2c 值: 我需要 in 和 if 语句的组合来表示数据帧,这对我有用。
sale_method = pd.DataFrame(model_data["Sale Method"].str.upper())
sale_method["sale_classification"] = np.where(
sale_method["Sale Method"].isin(["PRIVATE"]),
"private",
np.where(
sale_method["Sale Method"].str.contains("AUCTION"), "auction", "other"
),
)
为什么没有人谈论各种过滤方法的性能?事实上,这个话题经常在这里弹出(见示例)。我对一个大型数据集进行了自己的性能测试。这是非常有趣和有启发性的。
df = pd.DataFrame({'animals': np.random.choice(['cat', 'dog', 'mouse', 'birds'], size=10**7),
'number': np.random.randint(0,100, size=(10**7,))})
df.info()
<class 'pandas.core.frame.DataFrame'>
RangeIndex: 10000000 entries, 0 to 9999999
Data columns (total 2 columns):
# Column Dtype
--- ------ -----
0 animals object
1 number int64
dtypes: int64(1), object(1)
memory usage: 152.6+ MB
%%timeit
# .isin() by one column
conditions = ['cat', 'dog']
df[df.animals.isin(conditions)]
367 ms ± 2.34 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%%timeit
# .query() by one column
conditions = ['cat', 'dog']
df.query('animals in @conditions')
395 ms ± 3.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%%timeit
# .loc[]
df.loc[(df.animals=='cat')|(df.animals=='dog')]
987 ms ± 5.17 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%%timeit
df[df.apply(lambda x: x['animals'] in ['cat', 'dog'], axis=1)]
41.9 s ± 490 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%%timeit
new_df = df.set_index('animals')
new_df.loc[['cat', 'dog'], :]
3.64 s ± 62.5 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%%timeit
new_df = df.set_index('animals')
new_df[new_df.index.isin(['cat', 'dog'])]
469 ms ± 8.98 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
%%timeit
s = pd.Series(['cat', 'dog'], name='animals')
df.merge(s, on='animals', how='inner')
796 ms ± 30.9 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
因此,该方法被证明是最快的,而该方法是最慢的,这并不奇怪。isin
apply()
df.query('country.isin(@countries_to_keep).values')
# Or alternatively:
df.query('country.isin(["UK", "China"]).values')
若要否定查询,请使用:~
df.query('~country.isin(@countries_to_keep).values')
更新:
另一种方法是使用比较运算符:
df.query('country == @countries_to_keep')
# Or alternatively:
df.query('country == ["UK", "China"]')
要否定查询,请使用:!=
df.query('country != @countries_to_keep')
评论
groupby
可用于将数据帧“拆分”为两个
如果目标是将一个数据帧划分为两个数据帧,其中一个数据帧具有要保留的国家/地区,而另一个数据帧没有,则可以在调用中使用调用创建的布尔掩码将数据帧拆分为两个:have 和 have-not。isin
groupby
df = pd.DataFrame({'country': ['US', 'UK', 'Germany', 'China'], 'value': range(4)})
countries_to_keep = ['UK', 'China']
df1, df2 = [g for _, g in df.groupby(df['country'].isin(countries_to_keep))]
eval()
也可以使用
query()
建议在其他地方计算数值表达式。一个相关的方法是 。它可用于创建布尔掩码和过滤帧。它可以与可能在其他地方创建的其他蒙版一起使用,以实现更灵活的过滤。eval()
msk = df.eval('country in @countries_to_keep')
to_keep = df[msk] # in
not_keep = df[~msk] # not in
这很有用的一种特殊情况是,如果要使用条件筛选单个列,则内存效率非常低,因为它会创建筛选帧的副本,该副本需要再次针对单个列进行筛选,而使用布尔掩码列标签组合一次性选择该列。 可以做同样的事情。1query
loc
eval()
df = pd.DataFrame({'country': ['US', 'UK', 'Germany', 'China']*25000})
df[[f"col{i}" for i in range(50)]] = np.random.rand(100000, 50)
countries_to_keep = ['UK', 'China']
filtered = df.loc[df.eval('country==@countries_to_keep'), 'col1']
1 内存分析器测试:
import numpy as np
import pandas as pd
%load_ext memory_profiler
df = pd.DataFrame({'country': ['US', 'UK', 'Germany', 'China']*25000})
df[[f"col{i}" for i in range(50)]] = np.random.rand(100000, 50)
countries_to_keep = ['UK', 'China']
%memit x = df.loc[df.eval('country==@countries_to_keep'), 'col1']
# peak memory: 157.28 MiB, increment: 5.44 MiB
%memit y = df.query('country==@countries_to_keep')['col1']
# peak memory: 195.39 MiB, increment: 38.11 MiB
%memit z = df.loc[df['country'].isin(countries_to_keep), 'col1']
# peak memory: 176.93 MiB, increment: 0.76 MiB
评论
~