Grouping, splitting and picking top rows in a dataframe

问题

Problem

In the following dataframe df:

import random
import pandas as pd
random.seed(999)
sz = 50

qty = {'one': 1, 'two': 2, 'three': 3}

thing = (random.choice(['one', 'two', 'three']) for _ in range(sz))
order = (random.choice(['ascending', 'descending']) for _ in range(sz))
value = (random.randint(0, 100) for _ in range(sz))

df = pd.DataFrame({'thing': thing, 'order': order, 'value': value})

... I would like to:

1. Group by thing
2. Split it by order
3. Sort it by value for the thing per its order
4. Pick up the top qty for that thing

Expected Result

    thing       order  value
0     one   ascending     17
1     one  descending      1
2     two   ascending     28
3     two   ascending     30
4     two  descending     13
5     two  descending     38
6   three   ascending      6
7   three   ascending     27
8   three   ascending     35
9   three  descending      4
10  three  descending      5
11  three  descending      6

Manually coded to get the result by:

one_a = df[(df.thing == 'one') & (df.order == 'ascending')].reset_index(drop=True).sort_values('value', ascending='True').head(qty['one'])
one_d = df[(df.thing == 'one') & (df.order == 'descending')].reset_index(drop=True).sort_values('value', ascending='False').head(qty['one'])
two_a = df[(df.thing == 'two') & (df.order == 'ascending')].reset_index(drop=True).sort_values('value', ascending='True').head(qty['two'])
two_d = df[(df.thing == 'two') & (df.order == 'descending')].reset_index(drop=True).sort_values('value', ascending='False').head(qty['two'])
three_a = df[(df.thing == 'three') & (df.order == 'ascending')].reset_index(drop=True).sort_values('value', ascending='True').head(qty['three'])
three_d = df[(df.thing == 'three') & (df.order == 'descending')].reset_index(drop=True).sort_values('value', ascending='False').head(qty['three'])

print(pd.concat([one_a, one_d, two_a, two_d, three_a, three_d], ignore_index=True))

Question

Is it possible to achieve this using groupby, sort_values and set_index?

回答1:

One problem is to select ascending and descending separately. we can go around that by inverting descending:

df.loc[df.order=='descending','value']*= -1

s=(df.sort_values('value').groupby(['thing','order'])
     .cumcount()
     .reindex(df.index)
  )

out = df[s<df['thing'].map(qty)].sort_values(['thing','order'])
out.loc[out.order=='descending', 'value'] *= 1

Output:

    thing       order  value
14    one   ascending     17
27    one  descending      1
13  three   ascending      6
17  three   ascending     35
38  three   ascending     27
4   three  descending      5
23  three  descending      4
37  three  descending      6
21    two   ascending     28
42    two   ascending     30
6     two  descending     38
9     two  descending     13

来源：https://stackoverflow.com/questions/64864630/grouping-splitting-and-picking-top-rows-in-a-dataframe