我有一个数据帧,基本上是几个时间序列堆叠在一起。每个时间序列都有一个唯一的标签(组),它们具有不同的日期范围。
date = pd.to_datetime(pd.Series(['2010-01-01', '2010-01-02', '2010-01-03',
'2010-01-06', '2010-01-01', '2010-01-03']))
group = [1,1,1,1, 2, 2]
value = [1,2,3,4,5,6]
df = pd.DataFrame({'date':date, 'group':group, 'value':value})
df
date group value
0 2010-01-01 1 1
1 2010-01-02 1 2
2 2010-01-03 1 3
3 2010-01-06 1 4
4 2010-01-01 2 5
5 2010-01-03 2 6
我想重新采样数据,以便每个日期和组合的组合都有一个条目(如果当天没有观察或者它超出日期范围,则填充值为NaN)。示例输出将是:
date group value
2010-01-01 1 1
2010-01-02 1 2
2010-01-03 1 3
2010-01-04 1 NaN
2010-01-05 1 NaN
2010-01-06 1 4
2010-01-01 2 5
2010-01-02 2 NaN
2010-01-03 2 6
2010-01-04 2 NaN
2010-01-05 2 NaN
2010-01-06 2 NaN
我有一个有效的解决方案,但我怀疑有更好的方法。我的解决方案是首先转动数据然后取消堆叠,分组和重新取样。基本上所有真正需要的是进行分组和重新采样,但是使用整个日期列的最大值和最小值指定重采样的最大和最小范围,但我无论如何都看不到这样做。
df = (df.pivot(index='dates', columns='groups', values='values')
.unstack()
.reset_index()
.set_index('dates')
.groupby('groups').resample('D').asfreq()
.drop('groups', axis=1)
.reset_index()
.rename(columns={0:'values'}))[['dates','groups', 'values']]
答案 0 :(得分:3)
另一种方式:
import pandas as pd
from itertools import product
date = pd.to_datetime(pd.Series(['2010-01-01', '2010-01-02', '2010-01-03',
'2010-01-06', '2010-01-01', '2010-01-03']))
group = [1,1,1,1, 2, 2]
value = [1,2,3,4,5,6]
df = pd.DataFrame({'date':date, 'group':group, 'value':value})
dates = pd.date_range(df.date.min(), df.date.max())
groups = df.group.unique()
df = (pd.DataFrame(list(product(dates, groups)), columns=['date', 'group'])
.merge(df, on=['date', 'group'], how='left')
.sort_values(['group', 'date'])
.reset_index(drop=True))
df
# date group value
#0 2010-01-01 1 1.0
#1 2010-01-02 1 2.0
#2 2010-01-03 1 3.0
#3 2010-01-04 1 NaN
#4 2010-01-05 1 NaN
#5 2010-01-06 1 4.0
#6 2010-01-01 2 5.0
#7 2010-01-02 2 NaN
#8 2010-01-03 2 6.0
#9 2010-01-04 2 NaN
#10 2010-01-05 2 NaN
#11 2010-01-06 2 NaN
答案 1 :(得分:2)
归功于zipa以确保日期正确无误。我已经编辑了我的帖子来纠正我的错误。
设置索引然后使用pandas.MultiIndex.from_product
生成值的笛卡尔乘积。我还使用fill_value=0
填写缺少的值。
d = df.set_index(['date', 'group'])
midx = pd.MultiIndex.from_product(
[pd.date_range(df.date.min(), df.date.max()), df.group.unique()],
names=d.index.names
)
d.reindex(midx, fill_value=0).reset_index()
date group value
0 2010-01-01 1 1
1 2010-01-01 2 5
2 2010-01-02 1 2
3 2010-01-02 2 0
4 2010-01-03 1 3
5 2010-01-03 2 6
6 2010-01-04 1 0
7 2010-01-04 2 0
8 2010-01-05 1 0
9 2010-01-05 2 0
10 2010-01-06 1 4
11 2010-01-06 2 0
或者
d = df.set_index(['date', 'group'])
midx = pd.MultiIndex.from_product(
[pd.date_range(df.date.min(), df.date.max()), df.group.unique()],
names=d.index.names
)
d.reindex(midx).reset_index()
date group value
0 2010-01-01 1 1.0
1 2010-01-01 2 5.0
2 2010-01-02 1 2.0
3 2010-01-02 2 NaN
4 2010-01-03 1 3.0
5 2010-01-03 2 6.0
6 2010-01-04 1 NaN
7 2010-01-04 2 NaN
8 2010-01-05 1 NaN
9 2010-01-05 2 NaN
10 2010-01-06 1 4.0
11 2010-01-06 2 NaN
我们可以做的另一个舞蹈是OP尝试的清理版本。我再次使用fill_value=0
填写缺失值。我们可以将其留下来制作NaN
。
df.set_index(['date', 'group']) \
.unstack(fill_value=0) \
.asfreq('D', fill_value=0) \
.stack().reset_index()
date group value
0 2010-01-01 1 1
1 2010-01-01 2 5
2 2010-01-02 1 2
3 2010-01-02 2 0
4 2010-01-03 1 3
5 2010-01-03 2 6
6 2010-01-04 1 0
7 2010-01-04 2 0
8 2010-01-05 1 0
9 2010-01-05 2 0
10 2010-01-06 1 4
11 2010-01-06 2 0
或者
df.set_index(['date', 'group']) \
.unstack() \
.asfreq('D') \
.stack(dropna=False).reset_index()
date group value
0 2010-01-01 1 1.0
1 2010-01-01 2 5.0
2 2010-01-02 1 2.0
3 2010-01-02 2 NaN
4 2010-01-03 1 3.0
5 2010-01-03 2 6.0
6 2010-01-04 1 NaN
7 2010-01-04 2 NaN
8 2010-01-05 1 NaN
9 2010-01-05 2 NaN
10 2010-01-06 1 4.0
11 2010-01-06 2 NaN