数据科学和人工智能技术笔记 六、日期时间预处理

六、日期时间预处理

作者：Chris Albon

译者：飞龙

协议：CC BY-NC-SA 4.0

把日期和时间拆成多个特征

# 加载库
import pandas as pd

# 创建数据帧
df = pd.DataFrame()

# 创建五个日期
df['date'] = pd.date_range('1/1/2001', periods=150, freq='W')

# 为年月日，时分秒创建特征
df['year'] = df['date'].dt.year
df['month'] = df['date'].dt.month
df['day'] = df['date'].dt.day
df['hour'] = df['date'].dt.hour
df['minute'] = df['date'].dt.minute

# 展示三行
df.head(3)

	date	year	month	day	hour	minute
0	2001-01-07	2001	1	7	0	0
1	2001-01-14	2001	1	14	0	0
2	2001-01-21	2001	1	21	0	0

计算日期时间之间的差

# 加载库
import pandas as pd

# 创建数据帧
df = pd.DataFrame()

# 创建两个 datetime 特征
df['Arrived'] = [pd.Timestamp('01-01-2017'), pd.Timestamp('01-04-2017')]
df['Left'] = [pd.Timestamp('01-01-2017'), pd.Timestamp('01-06-2017')]

# 计算特征之间的间隔
df['Left'] - df['Arrived']

'''
0   0 days
1   2 days
dtype: timedelta64[ns] 
'''

# 计算特征之间的间隔
pd.Series(delta.days for delta in (df['Left'] - df['Arrived']))

'''
0    0
1    2
dtype: int64 
'''

将字符串转换为日期

# 加载库
import numpy as np
import pandas as pd

# 创建字符串
date_strings = np.array(['03-04-2005 11:35 PM',
                         '23-05-2010 12:01 AM',
                         '04-09-2009 09:09 PM'])

如果errors="coerce"那么任何问题都不会产生错误（默认行为），而是将导致错误的值设置为NaT（即缺失值）。

代码	描述	示例
%Y	整年	2001
%m	零填充的月份	04
%d	零填充的日期	09
%I	零填充的小时（12 小时）	02
%p	AM 或 PM	AM
%M	零填充的分钟	05
%S	零填充的秒钟	09

# 转换为 datetime
[pd.to_datetime(date, format="%d-%m-%Y %I:%M %p", errors="coerce") for date in date_strings]

'''
[Timestamp('2005-04-03 23:35:00'),
 Timestamp('2010-05-23 00:01:00'),
 Timestamp('2009-09-04 21:09:00')] 
'''

转换 pandas 列的时区

# 加载库
import pandas as pd
from pytz import all_timezones

# 展示十个时区
all_timezones[0:10]

'''
['Africa/Abidjan',
 'Africa/Accra',
 'Africa/Addis_Ababa',
 'Africa/Algiers',
 'Africa/Asmara',
 'Africa/Asmera',
 'Africa/Bamako',
 'Africa/Bangui',
 'Africa/Banjul',
 'Africa/Bissau'] 
'''

# 创建十个日期
dates = pd.Series(pd.date_range('2/2/2002', periods=10, freq='M'))

# 设置时区
dates_with_abidjan_time_zone = dates.dt.tz_localize('Africa/Abidjan')

# 查看 pandas 序列
dates_with_abidjan_time_zone

'''
0   2002-02-28 00:00:00+00:00
1   2002-03-31 00:00:00+00:00
2   2002-04-30 00:00:00+00:00
3   2002-05-31 00:00:00+00:00
4   2002-06-30 00:00:00+00:00
5   2002-07-31 00:00:00+00:00
6   2002-08-31 00:00:00+00:00
7   2002-09-30 00:00:00+00:00
8   2002-10-31 00:00:00+00:00
9   2002-11-30 00:00:00+00:00
dtype: datetime64[ns, Africa/Abidjan] 
'''

# 转换时区
dates_with_london_time_zone = dates_with_abidjan_time_zone.dt.tz_convert('Europe/London')

# 查看 pandas 序列
dates_with_london_time_zone

'''
0   2002-02-28 00:00:00+00:00
1   2002-03-31 00:00:00+00:00
2   2002-04-30 01:00:00+01:00
3   2002-05-31 01:00:00+01:00
4   2002-06-30 01:00:00+01:00
5   2002-07-31 01:00:00+01:00
6   2002-08-31 01:00:00+01:00
7   2002-09-30 01:00:00+01:00
8   2002-10-31 00:00:00+00:00
9   2002-11-30 00:00:00+00:00
dtype: datetime64[ns, Europe/London] 
'''

编码星期

# 加载库
import pandas as pd

# 创建数据集
dates = pd.Series(pd.date_range('2/2/2002', periods=3, freq='M'))

# 查看数据
dates

'''
0   2002-02-28
1   2002-03-31
2   2002-04-30
dtype: datetime64[ns] 
'''

# 查看星期
dates.dt.weekday_name

'''
0    Thursday
1      Sunday
2     Tuesday
dtype: object 
'''

处理时间序列中的缺失值

# 加载库
import pandas as pd
import numpy as np

# 创建日期
time_index = pd.date_range('01/01/2010', periods=5, freq='M')

# 创建数据帧，设置索引
df = pd.DataFrame(index=time_index)

# 创建带有一些缺失值的特征
df['Sales'] = [1.0,2.0,np.nan,np.nan,5.0]

# 对缺失值执行插值
df.interpolate()

	Sales
2010-01-31	1.0
2010-02-28	2.0
2010-03-31	3.0
2010-04-30	4.0
2010-05-31	5.0

# 前向填充
df.ffill()

	Sales
2010-01-31	1.0
2010-02-28	2.0
2010-03-31	2.0
2010-04-30	2.0
2010-05-31	5.0

# 后向填充
df.bfill()

	Sales
2010-01-31	1.0
2010-02-28	2.0
2010-03-31	5.0
2010-04-30	5.0
2010-05-31	5.0

# 对缺失值执行插值
df.interpolate(limit=1, limit_direction='forward')

	Sales
2010-01-31	1.0
2010-02-28	2.0
2010-03-31	3.0
2010-04-30	NaN
2010-05-31	5.0

处理时区

# 加载库
import pandas as pd
from pytz import all_timezones

# 展示十个时区
all_timezones[0:10]

'''
['Africa/Abidjan',
 'Africa/Accra',
 'Africa/Addis_Ababa',
 'Africa/Algiers',
 'Africa/Asmara',
 'Africa/Asmera',
 'Africa/Bamako',
 'Africa/Bangui',
 'Africa/Banjul',
 'Africa/Bissau'] 
'''

# 创建 datetime
pd.Timestamp('2017-05-01 06:00:00', tz='Europe/London')

# Timestamp('2017-05-01 06:00:00+0100', tz='Europe/London') 

# 创建 datetime
date = pd.Timestamp('2017-05-01 06:00:00')

# 设置时区
date_in_london = date.tz_localize('Europe/London')

# 修改时区
date_in_london.tz_convert('Africa/Abidjan')

# Timestamp('2017-05-01 05:00:00+0000', tz='Africa/Abidjan') 

平移时间特征

# 加载库
import pandas as pd

# 创建数据帧
df = pd.DataFrame()

# 创建数据
df['dates'] = pd.date_range('1/1/2001', periods=5, freq='D')
df['stock_price'] = [1.1,2.2,3.3,4.4,5.5]

# 将值平移一行
df['previous_days_stock_price'] = df['stock_price'].shift(1)

# 展示数据帧
df

	dates	stock_price	previous_days_stock_price
0	2001-01-01	1.1	NaN
1	2001-01-02	2.2	1.1
2	2001-01-03	3.3	2.2
3	2001-01-04	4.4	3.3
4	2001-01-05	5.5	4.4

滑动时间窗口

# 加载库
import pandas as pd

# 创建 datetime
time_index = pd.date_range('01/01/2010', periods=5, freq='M')

# 创建数据帧，设置索引
df = pd.DataFrame(index=time_index)

# 创建特征
df['Stock_Price'] = [1,2,3,4,5]

# 计算滑动均值
df.rolling(window=2).mean()

	Stock_Price
2010-01-31	NaN
2010-02-28	1.5
2010-03-31	2.5
2010-04-30	3.5
2010-05-31	4.5

# 识别滑动时间窗口中的最大值
df.rolling(window=2).max()

	Stock_Price
2010-01-31	NaN
2010-02-28	2.0
2010-03-31	3.0
2010-04-30	4.0
2010-05-31	5.0

选择日期时间范围

# 加载库
import pandas as pd

# 创建数据帧
df = pd.DataFrame()

# 创建 datetime
df['date'] = pd.date_range('1/1/2001', periods=100000, freq='H')

如果数据帧未按时间索引，请使用此方法。

# 选择两个日期时间之间的观测
df[(df['date'] > '2002-1-1 01:00:00') & (df['date'] <= '2002-1-1 04:00:00')]

	date
8762	2002-01-01 02:00:00
8763	2002-01-01 03:00:00
8764	2002-01-01 04:00:00

如果数据帧按时间索引，请使用此方法。

# 设置索引
df = df.set_index(df['date'])

# 选择两个日期时间之间的观测
df.loc['2002-1-1 01:00:00':'2002-1-1 04:00:00']

	date
date
2002-01-01 01:00:00	2002-01-01 01:00:00
2002-01-01 02:00:00	2002-01-01 02:00:00
2002-01-01 03:00:00	2002-01-01 03:00:00
2002-01-01 04:00:00	2002-01-01 04:00:00