问题
I have a dataframe as follows:
ref, type, amount
001, foo, 10
001, foo, 5
001, bar, 50
001, bar, 5
001, test, 100
001, test, 90
002, foo, 20
002, foo, 35
002, bar, 75
002, bar, 80
002, test, 150
002, test, 110
This is what I'm trying to get:
ref, type, amount, foo, bar, test
001, foo, 10, 15, 55, 190
001, foo, 5, 15, 55, 190
001, bar, 50, 15, 55, 190
001, bar, 5, 15, 55, 190
001, test, 100, 15, 55, 190
001, test, 90, 15, 55, 190
002, foo, 20, 55, 155, 260
002, foo, 35, 55, 155, 260
002, bar, 75, 55, 155, 260
002, bar, 80, 55, 155, 260
002, test, 150, 55, 155, 260
002, test, 110, 55, 155, 260
So I have this:
df.groupby('ref')['amount'].transform(sum)
But how can I filter it such that the above only applies to rows where type=foo or bar or test?
回答1:
A solution using pivot table :
>>> b = pd.pivot_table(df, values='amount', index=['ref'], columns=['type'], aggfunc=np.sum)
>>> b
type bar foo test
ref
1 55 15 190
2 155 55 260
>>> pd.merge(df, b, left_on='ref', right_index=True)
ref type amount bar foo test
0 1 foo 10 55 15 190
1 1 foo 5 55 15 190
2 1 bar 50 55 15 190
3 1 bar 5 55 15 190
4 1 test 100 55 15 190
5 1 test 90 55 15 190
6 2 foo 20 155 55 260
7 2 foo 35 155 55 260
8 2 bar 75 155 55 260
9 2 bar 80 155 55 260
10 2 test 150 155 55 260
11 2 test 110 155 55 260
回答2:
I think you need groupby with unstack and then merge to original DataFrame:
df1 = df.groupby(['ref','type'])['amount'].sum().unstack().reset_index()
print (df1)
type ref bar foo test
0 001 55 15 190
1 002 155 55 260
df = pd.merge(df, df1, on='ref')
print (df)
ref type amount sums bar foo test
0 001 foo 10 15 55 15 190
1 001 foo 5 15 55 15 190
2 001 bar 50 55 55 15 190
3 001 bar 5 55 55 15 190
4 001 test 100 190 55 15 190
5 001 test 90 190 55 15 190
6 002 foo 20 55 155 55 260
7 002 foo 35 55 155 55 260
8 002 bar 75 155 155 55 260
9 002 bar 80 155 155 55 260
10 002 test 150 260 155 55 260
11 002 test 110 260 155 55 260
Timings:
In [506]: %timeit (pd.merge(df, df.groupby(['ref','type'])['amount'].sum().unstack().reset_index(), on='ref'))
100 loops, best of 3: 3.4 ms per loop
In [507]: %timeit (pd.merge(df, pd.pivot_table(df, values='amount', index=['ref'], columns=['type'], aggfunc=np.sum), left_on='ref', right_index=True))
100 loops, best of 3: 4.99 ms per loop
来源：https://stackoverflow.com/questions/39750590/pandas-groupby-sum