pandas 将年月日拆分成高维 onehot 特征

如果是单变量时间序列，已知的只有 datetime 信息，怎么从中提取特征呢？

可以将一个日期拆分出多维信息，比如：是否是周末？是上旬中旬还是下旬？是否是节假日？是春是夏是秋还是冬？

人类活动和这些因素有着很强的联系，但是如果只将时间编码成一个日期的字符串，机器学习算法是无法提取到有效信息的。

下面就来将上图中的 datetime 列拆分成高维特征

def half_month(day):
    return 1 if day < 15 else 2

temp.loc[:,'year'] = temp['datetime'].apply(lambda x:x.year)
temp.loc[:,'month'] = temp['datetime'].apply(lambda x:x.month)
temp.loc[:,'day'] = temp['datetime'].apply(lambda x: x.day)
temp.loc[:,'day_of_week'] = temp['datetime'].apply(lambda x: x.dayofweek)
temp.loc[:,'weekend'] = temp['datetime'].apply(lambda x:x.dayofweek>4)
temp.loc[:,'half_month'] = temp['datetime'].apply(lambda x: half_month(x.day))

temp.columns
'''
Index(['datetime', 'TEMP', 'year', 'month', 'day', 'day_of_week', 'weekend', 'half_month'],
      dtype='object')
'''
temp.head()

使用 pandas.get_dummies 函数把指定列编码成 onehot 形式

columns2onehot = ['year','month','day','day_of_week','half_month']
onehot = pd.get_dummies(temp, columns=columns2onehot)
onehot.columns
'''
Index(['datetime', 'TEMP', 'weekend', 'year_2010', 'year_2011', 'year_2012',
       'year_2013', 'year_2014', 'month_1', 'month_2', 'month_3', 'month_4',
       'month_5', 'month_6', 'month_7', 'month_8', 'month_9', 'month_10',
       'month_11', 'month_12', 'day_1', 'day_2', 'day_3', 'day_4', 'day_5',
       'day_6', 'day_7', 'day_8', 'day_9', 'day_10', 'day_11', 'day_12',
       'day_13', 'day_14', 'day_15', 'day_16', 'day_17', 'day_18', 'day_19',
       'day_20', 'day_21', 'day_22', 'day_23', 'day_24', 'day_25', 'day_26',
       'day_27', 'day_28', 'day_29', 'day_30', 'day_31', 'day_of_week_0',
       'day_of_week_1', 'day_of_week_2', 'day_of_week_3', 'day_of_week_4',
       'day_of_week_5', 'day_of_week_6', 'half_month_1', 'half_month_2'],
      dtype='object')
'''
onehot.head()

datetime	TEMP	weekend	half_month	year_2010	year_2011	year_2012	year_2013	year_2014	month_1	…	day_29	day_30	day_31	day_of_week_0	day_of_week_1	day_of_week_2	day_of_week_3	day_of_week_4	day_of_week_5	day_of_week_6
2010-01-01 00:00:00	-11.0	False	1	1	0	0	0	0	1	…	0	0	0	0	0	0	0	1	0	0
2010-01-01 01:00:00	-12.0	False	1	1	0	0	0	0	1	…	0	0	0	0	0	0	0	1	0	0
2010-01-01 02:00:00	-11.0	False	1	1	0	0	0	0	1	…	0	0	0	0	0	0	0	1	0	0
2010-01-01 03:00:00	-14.0	False	1	1	0	0	0	0	1	…	0	0	0	0	0	0	0	1	0	0
2010-01-01 04:00:00	-12.0	False	1	1	0	0	0	0	1	…	0	0	0	0	0	0	0	1	0	0

然后就可以将原来非独热编码的列删掉，就可以用机器学习模型训练了！