从 pandas dataframe 创建滑动 window 组

Question

I'm trying to pre-process my data for a ML regression problem.
With the following (simplified) data frame:

   grp day  score
0    A   1      2
1    A   1      4
2    A   2      6
3    A   2      8
4    A   3     10
5    A   3     12
6    A   4     14
7    A   4     16
8    A   5     18
9    A   5     20
10   B   1      2
11   B   2      4
12   B   3      8
13   B   4     16
14   B   5     32

I'm trying to create a list of 'sliding window' sequences based on the day column, so if I have X days, the first 2 days will have a target of the score Y days ahed.

In the example bellow I have 5 days in each group and for each 2 days I'm looking at the target of 2 days ahead, stopping when I've reached the end of the data frame:

So for example here are the 2 first groups for group A:

   grp day  score   target
0    A   1      2    16
1    A   1      4    16
2    A   2      6    16
3    A   2      8    16 <- last score value of day 4 (group A)

   grp day  score   target
0    A   2      6    20
1    A   2      8    20
2    A   3      10   20
3    A   3      12   20 <- last score value of day 5 (group A)

And for group B:

   grp day  score   target
10   B   1      2    16
11   B   2      4    16 <- last score value of day 4 (group B)

   grp day  score   target
10   B   2      4    32
11   B   3      8    32 <- last score value of day 5 (group B)

I've use factorize to get the days index and group like so:

groups = df.groupby(['grp'])
for _,grp in groups:
  days_row_index = grp['day'].factorize()[0]
  days_group = grp.groupby(days_row_index)
  ...

But I'm a bit lost... any help would be appreciated Update:

I've written the following clumsy code, to get me going... how can I improve it?

import pandas as pd
df = pd.DataFrame({'grp':['A','A','A','A','A','A','A','A','A','A','B','B','B','B','B'],
                   'day':['1','1','2','2','3','3','4','4','5','5','1','2','3','4','5'],
                   'score':[2,4,6,8,10,12,14,16,18,20,2,4,8,16,32]
                   })

print(df.head(15))

df2 = pd.DataFrame({'grp':[],
                    'day':[],
                    'score':[]})

groups = df.groupby(['grp'])
GROUP_SIZE = 2
LOOK_AHEAD = 2
sequences = []

for _,grp in groups:
  days_row_index = grp['day'].factorize()[0]
  days_group = grp.groupby(days_row_index)
  for _,day in days_group:
    day_index = int(day['day'].values[0])
    if day_index + LOOK_AHEAD < len(days_group):
      target = days_group.get_group(day_index + LOOK_AHEAD)['score'].values[-1]
      print(day_index,day_index + LOOK_AHEAD,day['score'].values[-1],"----------->",target)
      day['target'] = target
      df2 = pd.concat([df2,day])
      for i in range(0, GROUP_SIZE-1):
        if day_index + i >= len(days_group):
          break
        next_day = days_group.get_group(day_index + i)
        next_day['target'] = target
        df2 = pd.concat([df2,next_day])
      sequences.append(df2.copy())
      df2 = df2.iloc[0:0]
sequences

Answer 1

Building upon your proposed solution I wrote this little piece of code, which I'm positive it can be optimized so feel free anyone to improve it. Let me know if this was what you were looking for (I took the liberty to create another "mixed" group 'C' to test a more generalized approach).

import pandas as pd

# Create test dataframe
df = [
     ['A', 1, 2],
     ['A', 1, 4],
     ['A', 2, 6],
     ['A', 2, 8],
     ['A', 3, 10],
     ['A', 3, 12],
     ['A', 4, 14],
     ['A', 4, 16],
     ['A', 5, 18],
     ['A', 5, 20],
     ['B', 1, 2],
     ['B', 2, 4],
     ['B', 3, 8],
     ['B', 4, 16],
     ['B', 5, 32],
     ['C', 1, 2],
     ['C', 1, 4],
     ['C', 2, 8],
     ['C', 3, 16],
     ['C', 3, 20],
     ['C', 4, 24],
     ['C', 5, 28]
     ]
df = pd.DataFrame(df, columns = ['grp', 'day', 'score'])

# Processing
groups = df.groupby(['grp'])
for _,grp in groups:
  days_row_index = grp['day'].factorize()[0]
  i = min(days_row_index)
  while i < max(days_row_index) - 2:
      idx = (days_row_index == i) | (days_row_index == i + 1)
      # Create list of targets for every subgroup
      print([grp['score'].values[days_row_index == i + 3][-1]]*sum(idx))
      i += 1