python - Pandas 前一组最小/最大

Question

在 Pandas 中，我有这样的数据集:

                     Value
2005-08-03 23:15:00   10.5
2005-08-03 23:30:00   10.0
2005-08-03 23:45:00   10.0
2005-08-04 00:00:00   10.5
2005-08-04 00:15:00   10.5
2005-08-04 00:30:00   11.0
2005-08-04 00:45:00   10.5
2005-08-04 01:00:00   11.0
...
2005-08-04 23:15:00   14.0
2005-08-04 23:30:00   13.5
2005-08-04 23:45:00   13.0
2005-08-05 00:00:00   13.5
2005-08-05 00:15:00   14.0
2005-08-05 00:30:00   14.0
2005-08-05 00:45:00   14.5

首先，我想按日期对数据进行分组并将每个组的最大值存储在新列中，我使用以下代码来完成此任务:

df['ValueMaxInGroup'] = df.groupby(pd.TimeGrouper('D'))['Value'].transform(max)

现在我想创建另一列来存储前一组最大值，因此所需的数据框如下所示:

                     Value  ValueMaxInGroup  ValueMaxInPrevGroup
2005-08-03 23:15:00   10.5             10.5                  NaN
2005-08-03 23:30:00   10.0             10.5                  NaN
2005-08-03 23:45:00   10.0             10.5                  NaN
2005-08-04 00:00:00   10.5             14.0                 10.5
2005-08-04 00:15:00   10.5             14.0                 10.5
2005-08-04 00:30:00   11.0             14.0                 10.5
2005-08-04 00:45:00   10.5             14.0                 10.5
2005-08-04 01:00:00   11.0             14.0                 10.5
...
2005-08-04 23:15:00   14.0             14.0                 10.5
2005-08-04 23:30:00   13.5             14.0                 10.5
2005-08-04 23:45:00   13.0             14.0                 10.5
2005-08-05 00:00:00   13.5             14.5                 14.0
2005-08-05 00:15:00   14.0             14.5                 14.0
2005-08-05 00:30:00   14.0             14.5                 14.0
2005-08-05 00:45:00   14.5             14.5                 14.0

因此，为了简单地获取上一行的值，我使用了

df['ValueInPrevRow'] = df.shift(1)['Value']

有没有办法获得另一组的最小/最大/f(x)？我以为

df['ValueMaxInPrevGroup'] = df.groupby(pd.TimeGrouper('D')).shift(1)['Value'].transform(max)

但是没有成功。

Answer 1

您可以使用groupby/agg、shift和merge获得所需的结果:

import numpy as np
import pandas as pd
df = pd.DataFrame({'Value': [10.5, 10.0, 10.0, 10.5, 10.5, 11.0, 10.5, 11.0, 14.0, 13.5, 13.0, 13.5, 14.0, 14.0, 14.5]}, index=['2005-08-03 23:15:00', '2005-08-03 23:30:00', '2005-08-03 23:45:00', '2005-08-04 00:00:00', '2005-08-04 00:15:00', '2005-08-04 00:30:00', '2005-08-04 00:45:00', '2005-08-04 01:00:00', '2005-08-04 23:15:00', '2005-08-04 23:30:00', '2005-08-04 23:45:00', '2005-08-05 00:00:00', '2005-08-05 00:15:00', '2005-08-05 00:30:00', '2005-08-05 00:45:00']) 
df.index = pd.DatetimeIndex(df.index)

# This is equivalent to
# df['group'] = pd.to_datetime(df.index.date)
# when freq='D', but the version below works with any freq string, not just `'D'`.
grouped = df.groupby(pd.TimeGrouper('D'))
labels, uniqs, ngroups = grouped.grouper.group_info
df['group'] = grouped.grouper.binlabels[labels]

result = grouped[['Value']].agg(max)
result = result.rename(columns={'Value':'Max'})
result['PreviouMax'] = result['Max'].shift(1)

df = pd.merge(df, result, left_on=['group'], right_index=True)
print(df)

产量

                     Value      group   Max  PreviouMax
2005-08-03 23:15:00   10.5 2005-08-03  10.5         NaN
2005-08-03 23:30:00   10.0 2005-08-03  10.5         NaN
2005-08-03 23:45:00   10.0 2005-08-03  10.5         NaN
2005-08-04 00:00:00   10.5 2005-08-04  14.0        10.5
2005-08-04 00:15:00   10.5 2005-08-04  14.0        10.5
2005-08-04 00:30:00   11.0 2005-08-04  14.0        10.5
2005-08-04 00:45:00   10.5 2005-08-04  14.0        10.5
2005-08-04 01:00:00   11.0 2005-08-04  14.0        10.5
2005-08-04 23:15:00   14.0 2005-08-04  14.0        10.5
2005-08-04 23:30:00   13.5 2005-08-04  14.0        10.5
2005-08-04 23:45:00   13.0 2005-08-04  14.0        10.5
2005-08-05 00:00:00   13.5 2005-08-05  14.5        14.0
2005-08-05 00:15:00   14.0 2005-08-05  14.5        14.0
2005-08-05 00:30:00   14.0 2005-08-05  14.5        14.0
2005-08-05 00:45:00   14.5 2005-08-05  14.5        14.0

<小时/>

这里的主要思想是使用groupby/agg而不是groupby/transform，这样我们就可以获得

result = grouped[['Value']].agg(max)
result = result.rename(columns={'Value':'Max'})
result['PreviouMax'] = result['Max'].shift(1)
#              Max  PreviouMax
# group                       
# 2005-08-03  10.5         NaN
# 2005-08-04  14.0        10.5
# 2005-08-05  14.5        14.0

然后所需的DataFrame可以表示为将df与 组日期的结果。