Pandas groupby: How to get a union of strings
Question:
I have a dataframe like this:
A B C
0 1 0.749065 This
1 2 0.301084 is
2 3 0.463468 a
3 4 0.643961 random
4 1 0.866521 string
5 2 0.120737 !
Calling
In [10]: print df.groupby("A")["B"].sum()
will return
A
1 1.615586
2 0.421821
3 0.463468
4 0.643961
Now I would like to do “the same” for column “C”. Because that column contains strings, sum() doesn’t work (although you might think that it would concatenate the strings). What I would really like to see is a list or set of the strings for each group, i.e.
A
1 {This, string}
2 {is, !}
3 {a}
4 {random}
I have been trying to find ways to do this.
Series.unique() (http://pandas.pydata.org/pandas-docs/stable/generated/pandas.Series.unique.html) doesn’t work, although
df.groupby("A")["B"]
is a
pandas.core.groupby.SeriesGroupBy object
so I was hoping any Series method would work. Any ideas?
Answers:
In [4]: df = read_csv(StringIO(data),sep='s+')
In [5]: df
Out[5]:
A B C
0 1 0.749065 This
1 2 0.301084 is
2 3 0.463468 a
3 4 0.643961 random
4 1 0.866521 string
5 2 0.120737 !
In [6]: df.dtypes
Out[6]:
A int64
B float64
C object
dtype: object
When you apply your own function, there is not automatic exclusions of non-numeric columns. This is slower, though, than the application of .sum() to the groupby
In [8]: df.groupby('A').apply(lambda x: x.sum())
Out[8]:
A B C
A
1 2 1.615586 Thisstring
2 4 0.421821 is!
3 3 0.463468 a
4 4 0.643961 random
sum by default concatenates
In [9]: df.groupby('A')['C'].apply(lambda x: x.sum())
Out[9]:
A
1 Thisstring
2 is!
3 a
4 random
dtype: object
You can do pretty much what you want
In [11]: df.groupby('A')['C'].apply(lambda x: "{%s}" % ', '.join(x))
Out[11]:
A
1 {This, string}
2 {is, !}
3 {a}
4 {random}
dtype: object
Doing this on a whole frame, one group at a time. Key is to return a Series
def f(x):
return Series(dict(A = x['A'].sum(),
B = x['B'].sum(),
C = "{%s}" % ', '.join(x['C'])))
In [14]: df.groupby('A').apply(f)
Out[14]:
A B C
A
1 2 1.615586 {This, string}
2 4 0.421821 {is, !}
3 3 0.463468 {a}
4 4 0.643961 {random}
You can use the apply method to apply an arbitrary function to the grouped data. So if you want a set, apply set. If you want a list, apply list.
>>> d
A B
0 1 This
1 2 is
2 3 a
3 4 random
4 1 string
5 2 !
>>> d.groupby('A')['B'].apply(list)
A
1 [This, string]
2 [is, !]
3 [a]
4 [random]
dtype: object
If you want something else, just write a function that does what you want and then apply that.
You may be able to use the aggregate (or agg) function to concatenate the values. (Untested code)
df.groupby('A')['B'].agg(lambda col: ''.join(col))
a simple solution would be :
>>> df.groupby(['A','B']).c.unique().reset_index()
You could try this:
df.groupby('A').agg({'B':'sum','C':'-'.join})
If you’d like to overwrite column B in the dataframe, this should work:
df = df.groupby('A',as_index=False).agg(lambda x:'n'.join(x))
Named aggregations with pandas >= 0.25.0
Since pandas version 0.25.0 we have named aggregations where we can groupby, aggregate and at the same time assign new names to our columns. This way we won’t get the MultiIndex columns, and the column names make more sense given the data they contain:
aggregate and get a list of strings
grp = df.groupby('A').agg(B_sum=('B','sum'),
C=('C', list)).reset_index()
print(grp)
A B_sum C
0 1 1.615586 [This, string]
1 2 0.421821 [is, !]
2 3 0.463468 [a]
3 4 0.643961 [random]
aggregate and join the strings
grp = df.groupby('A').agg(B_sum=('B','sum'),
C=('C', ', '.join)).reset_index()
print(grp)
A B_sum C
0 1 1.615586 This, string
1 2 0.421821 is, !
2 3 0.463468 a
3 4 0.643961 random
Following @Erfan’s good answer, most of the times in an analysis of aggregate values you want the unique possible combinations of these existing character values:
unique_chars = lambda x: ', '.join(x.unique())
(df
.groupby(['A'])
.agg({'C': unique_chars}))
I have a dataframe like this:
A B C
0 1 0.749065 This
1 2 0.301084 is
2 3 0.463468 a
3 4 0.643961 random
4 1 0.866521 string
5 2 0.120737 !
Calling
In [10]: print df.groupby("A")["B"].sum()
will return
A
1 1.615586
2 0.421821
3 0.463468
4 0.643961
Now I would like to do “the same” for column “C”. Because that column contains strings, sum() doesn’t work (although you might think that it would concatenate the strings). What I would really like to see is a list or set of the strings for each group, i.e.
A
1 {This, string}
2 {is, !}
3 {a}
4 {random}
I have been trying to find ways to do this.
Series.unique() (http://pandas.pydata.org/pandas-docs/stable/generated/pandas.Series.unique.html) doesn’t work, although
df.groupby("A")["B"]
is a
pandas.core.groupby.SeriesGroupBy object
so I was hoping any Series method would work. Any ideas?
In [4]: df = read_csv(StringIO(data),sep='s+')
In [5]: df
Out[5]:
A B C
0 1 0.749065 This
1 2 0.301084 is
2 3 0.463468 a
3 4 0.643961 random
4 1 0.866521 string
5 2 0.120737 !
In [6]: df.dtypes
Out[6]:
A int64
B float64
C object
dtype: object
When you apply your own function, there is not automatic exclusions of non-numeric columns. This is slower, though, than the application of .sum() to the groupby
In [8]: df.groupby('A').apply(lambda x: x.sum())
Out[8]:
A B C
A
1 2 1.615586 Thisstring
2 4 0.421821 is!
3 3 0.463468 a
4 4 0.643961 random
sum by default concatenates
In [9]: df.groupby('A')['C'].apply(lambda x: x.sum())
Out[9]:
A
1 Thisstring
2 is!
3 a
4 random
dtype: object
You can do pretty much what you want
In [11]: df.groupby('A')['C'].apply(lambda x: "{%s}" % ', '.join(x))
Out[11]:
A
1 {This, string}
2 {is, !}
3 {a}
4 {random}
dtype: object
Doing this on a whole frame, one group at a time. Key is to return a Series
def f(x):
return Series(dict(A = x['A'].sum(),
B = x['B'].sum(),
C = "{%s}" % ', '.join(x['C'])))
In [14]: df.groupby('A').apply(f)
Out[14]:
A B C
A
1 2 1.615586 {This, string}
2 4 0.421821 {is, !}
3 3 0.463468 {a}
4 4 0.643961 {random}
You can use the apply method to apply an arbitrary function to the grouped data. So if you want a set, apply set. If you want a list, apply list.
>>> d
A B
0 1 This
1 2 is
2 3 a
3 4 random
4 1 string
5 2 !
>>> d.groupby('A')['B'].apply(list)
A
1 [This, string]
2 [is, !]
3 [a]
4 [random]
dtype: object
If you want something else, just write a function that does what you want and then apply that.
You may be able to use the aggregate (or agg) function to concatenate the values. (Untested code)
df.groupby('A')['B'].agg(lambda col: ''.join(col))
a simple solution would be :
>>> df.groupby(['A','B']).c.unique().reset_index()
You could try this:
df.groupby('A').agg({'B':'sum','C':'-'.join})
If you’d like to overwrite column B in the dataframe, this should work:
df = df.groupby('A',as_index=False).agg(lambda x:'n'.join(x))
Named aggregations with pandas >= 0.25.0
Since pandas version 0.25.0 we have named aggregations where we can groupby, aggregate and at the same time assign new names to our columns. This way we won’t get the MultiIndex columns, and the column names make more sense given the data they contain:
aggregate and get a list of strings
grp = df.groupby('A').agg(B_sum=('B','sum'),
C=('C', list)).reset_index()
print(grp)
A B_sum C
0 1 1.615586 [This, string]
1 2 0.421821 [is, !]
2 3 0.463468 [a]
3 4 0.643961 [random]
aggregate and join the strings
grp = df.groupby('A').agg(B_sum=('B','sum'),
C=('C', ', '.join)).reset_index()
print(grp)
A B_sum C
0 1 1.615586 This, string
1 2 0.421821 is, !
2 3 0.463468 a
3 4 0.643961 random
Following @Erfan’s good answer, most of the times in an analysis of aggregate values you want the unique possible combinations of these existing character values:
unique_chars = lambda x: ', '.join(x.unique())
(df
.groupby(['A'])
.agg({'C': unique_chars}))