SQL-like window functions in PANDAS: Row Numbering in Python Pandas Dataframe
Question:
I come from a sql background and I use the following data processing step frequently:
- Partition the table of data by one or more fields
- For each partition, add a rownumber to each of its rows that ranks the row by one or more other fields, where the analyst specifies ascending or descending
EX:
df = pd.DataFrame({'key1' : ['a','a','a','b','a'],
'data1' : [1,2,2,3,3],
'data2' : [1,10,2,3,30]})
df
data1 data2 key1
0 1 1 a
1 2 10 a
2 2 2 a
3 3 3 b
4 3 30 a
I’m looking for how to do the PANDAS equivalent to this sql window function:
RN = ROW_NUMBER() OVER (PARTITION BY Key1 ORDER BY Data1 ASC, Data2 DESC)
data1 data2 key1 RN
0 1 1 a 1
1 2 10 a 2
2 2 2 a 3
3 3 3 b 1
4 3 30 a 4
I’ve tried the following which I’ve gotten to work where there are no ‘partitions’:
def row_number(frame,orderby_columns, orderby_direction,name):
frame.sort_index(by = orderby_columns, ascending = orderby_direction, inplace = True)
frame[name] = list(xrange(len(frame.index)))
I tried to extend this idea to work with partitions (groups in pandas) but the following didn’t work:
df1 = df.groupby('key1').apply(lambda t: t.sort_index(by=['data1', 'data2'], ascending=[True, False], inplace = True)).reset_index()
def nf(x):
x['rn'] = list(xrange(len(x.index)))
df1['rn1'] = df1.groupby('key1').apply(nf)
But I just got a lot of NaNs when I do this.
Ideally, there’d be a succinct way to replicate the window function capability of sql (i’ve figured out the window based aggregates…that’s a one liner in pandas)…can someone share with me the most idiomatic way to number rows like this in PANDAS?
Answers:
You can do this by using groupby
twice along with the rank
method:
In [11]: g = df.groupby('key1')
Use the min method argument to give values which share the same data1 the same RN:
In [12]: g['data1'].rank(method='min')
Out[12]:
0 1
1 2
2 2
3 1
4 4
dtype: float64
In [13]: df['RN'] = g['data1'].rank(method='min')
And then groupby these results and add the rank with respect to data2:
In [14]: g1 = df.groupby(['key1', 'RN'])
In [15]: g1['data2'].rank(ascending=False) - 1
Out[15]:
0 0
1 0
2 1
3 0
4 0
dtype: float64
In [16]: df['RN'] += g1['data2'].rank(ascending=False) - 1
In [17]: df
Out[17]:
data1 data2 key1 RN
0 1 1 a 1
1 2 10 a 2
2 2 2 a 3
3 3 3 b 1
4 3 30 a 4
It feels like there ought to be a native way to do this (there may well be!…).
pandas.lib.fast_zip()
can create a tuple array from a list of array. You can use this function to create a tuple series, and then rank it:
values = {'key1' : ['a','a','a','b','a','b'],
'data1' : [1,2,2,3,3,3],
'data2' : [1,10,2,3,30,20]}
df = pd.DataFrame(values, index=list("abcdef"))
def rank_multi_columns(df, cols, **kw):
data = []
for col in cols:
if col.startswith("-"):
flag = -1
col = col[1:]
else:
flag = 1
data.append(flag*df[col])
values = pd.lib.fast_zip(data)
s = pd.Series(values, index=df.index)
return s.rank(**kw)
rank = df.groupby("key1").apply(lambda df:rank_multi_columns(df, ["data1", "-data2"]))
print rank
the result:
a 1
b 2
c 3
d 2
e 4
f 1
dtype: float64
you can also use sort_values()
, groupby()
and finally cumcount() + 1
:
df['RN'] = df.sort_values(['data1','data2'], ascending=[True,False])
.groupby(['key1'])
.cumcount() + 1
print(df)
yields:
data1 data2 key1 RN
0 1 1 a 1
1 2 10 a 2
2 2 2 a 3
3 3 3 b 1
4 3 30 a 4
PS tested with pandas 0.18
Use groupby.rank function.
Here the working example.
df = pd.DataFrame({'C1':['a', 'a', 'a', 'b', 'b'], 'C2': [1, 2, 3, 4, 5]})
df
C1 C2
a 1
a 2
a 3
b 4
b 5
df["RANK"] = df.groupby("C1")["C2"].rank(method="first", ascending=True)
df
C1 C2 RANK
a 1 1
a 2 2
a 3 3
b 4 1
b 5 2
I come from a sql background and I use the following data processing step frequently:
- Partition the table of data by one or more fields
- For each partition, add a rownumber to each of its rows that ranks the row by one or more other fields, where the analyst specifies ascending or descending
EX:
df = pd.DataFrame({'key1' : ['a','a','a','b','a'],
'data1' : [1,2,2,3,3],
'data2' : [1,10,2,3,30]})
df
data1 data2 key1
0 1 1 a
1 2 10 a
2 2 2 a
3 3 3 b
4 3 30 a
I’m looking for how to do the PANDAS equivalent to this sql window function:
RN = ROW_NUMBER() OVER (PARTITION BY Key1 ORDER BY Data1 ASC, Data2 DESC)
data1 data2 key1 RN
0 1 1 a 1
1 2 10 a 2
2 2 2 a 3
3 3 3 b 1
4 3 30 a 4
I’ve tried the following which I’ve gotten to work where there are no ‘partitions’:
def row_number(frame,orderby_columns, orderby_direction,name):
frame.sort_index(by = orderby_columns, ascending = orderby_direction, inplace = True)
frame[name] = list(xrange(len(frame.index)))
I tried to extend this idea to work with partitions (groups in pandas) but the following didn’t work:
df1 = df.groupby('key1').apply(lambda t: t.sort_index(by=['data1', 'data2'], ascending=[True, False], inplace = True)).reset_index()
def nf(x):
x['rn'] = list(xrange(len(x.index)))
df1['rn1'] = df1.groupby('key1').apply(nf)
But I just got a lot of NaNs when I do this.
Ideally, there’d be a succinct way to replicate the window function capability of sql (i’ve figured out the window based aggregates…that’s a one liner in pandas)…can someone share with me the most idiomatic way to number rows like this in PANDAS?
You can do this by using groupby
twice along with the rank
method:
In [11]: g = df.groupby('key1')
Use the min method argument to give values which share the same data1 the same RN:
In [12]: g['data1'].rank(method='min')
Out[12]:
0 1
1 2
2 2
3 1
4 4
dtype: float64
In [13]: df['RN'] = g['data1'].rank(method='min')
And then groupby these results and add the rank with respect to data2:
In [14]: g1 = df.groupby(['key1', 'RN'])
In [15]: g1['data2'].rank(ascending=False) - 1
Out[15]:
0 0
1 0
2 1
3 0
4 0
dtype: float64
In [16]: df['RN'] += g1['data2'].rank(ascending=False) - 1
In [17]: df
Out[17]:
data1 data2 key1 RN
0 1 1 a 1
1 2 10 a 2
2 2 2 a 3
3 3 3 b 1
4 3 30 a 4
It feels like there ought to be a native way to do this (there may well be!…).
pandas.lib.fast_zip()
can create a tuple array from a list of array. You can use this function to create a tuple series, and then rank it:
values = {'key1' : ['a','a','a','b','a','b'],
'data1' : [1,2,2,3,3,3],
'data2' : [1,10,2,3,30,20]}
df = pd.DataFrame(values, index=list("abcdef"))
def rank_multi_columns(df, cols, **kw):
data = []
for col in cols:
if col.startswith("-"):
flag = -1
col = col[1:]
else:
flag = 1
data.append(flag*df[col])
values = pd.lib.fast_zip(data)
s = pd.Series(values, index=df.index)
return s.rank(**kw)
rank = df.groupby("key1").apply(lambda df:rank_multi_columns(df, ["data1", "-data2"]))
print rank
the result:
a 1
b 2
c 3
d 2
e 4
f 1
dtype: float64
you can also use sort_values()
, groupby()
and finally cumcount() + 1
:
df['RN'] = df.sort_values(['data1','data2'], ascending=[True,False])
.groupby(['key1'])
.cumcount() + 1
print(df)
yields:
data1 data2 key1 RN
0 1 1 a 1
1 2 10 a 2
2 2 2 a 3
3 3 3 b 1
4 3 30 a 4
PS tested with pandas 0.18
Use groupby.rank function.
Here the working example.
df = pd.DataFrame({'C1':['a', 'a', 'a', 'b', 'b'], 'C2': [1, 2, 3, 4, 5]})
df
C1 C2
a 1
a 2
a 3
b 4
b 5
df["RANK"] = df.groupby("C1")["C2"].rank(method="first", ascending=True)
df
C1 C2 RANK
a 1 1
a 2 2
a 3 3
b 4 1
b 5 2