计算数据帧列的最快方法

Question

I'm getting a pandas issue that I need help with.

On the one hand, I have a DataFrame that looks like the following:

   contributor_id     timestamp     edits    upper_month   lower_month
0      8             2018-01-01       1      2018-04-01    2018-02-01
1      26424341      2018-01-01       11     2018-04-01    2018-02-01
10     26870381      2018-01-01       465    2018-04-01    2018-02-01
22     28109145      2018-03-01       17     2018-06-01    2018-04-01
23     32769624      2018-01-01       84     2018-04-01    2018-02-01
25     32794352      2018-01-01       4      2018-04-01    2018-02-01

On the other hand, I have (available in another DF), a given index of dates:

2018-01-01, 2018-02-01, 2018-03-01, 2018-04-01, 2018-05-01, 2018-06-01, 2018-07-01, 2018-08-01, 2018-09-01, 2018-10-01, 2018-11-01, 2018-12-01.

I need to create a pd.Series that has as an index the previously shown index. The data of the pd.Series must be, for each date in the index:

if date >= lower_month and date <= upper_month, then I add a 1.

The objective is to count, per each date, the number of times that the date is between the upper and lower month values in the previous DataFrame.

The sample output pd.Series for this case would be:

2018-01-01    0
2018-02-01    5
2018-03-01    5
2018-04-01    6
2018-05-01    1
2018-06-01    1
2018-07-01    0
2018-08-01    0
2018-09-01    0
2018-10-01    0
2018-11-01    0
2018-12-01    0

Is there a fast way of doing this calculation, avoiding to traverse the first dataframe a huge amount of times?

Thank you all.

Answer 1

Use list comprehension with flattening for test membership between zipped columns converted to tuples and values in range, create DataFrame and sum in generator:

rng = pd.date_range('2018-01-01', freq='MS', periods=12)
vals = list(zip(df['lower_month'], df['upper_month']))

s = pd.Series({y: sum(y >= x1 and y <= x2 for x1, x2 in vals) for y in rng})

EDIT:

For better performance use count method, thank you @Stef:

s = pd.Series({y: [y >= x1 and y <= x2 for x1, x2 in vals].count(True) for y in rng})

print (s)
2018-01-01    0
2018-02-01    5
2018-03-01    5
2018-04-01    6
2018-05-01    1
2018-06-01    1
2018-07-01    0
2018-08-01    0
2018-09-01    0
2018-10-01    0
2018-11-01    0
2018-12-01    0
dtype: int64

Performace :

np.random.seed(123)

def random_dates(start, end, n=10000):

    start_u = start.value//10**9
    end_u = end.value//10**9

    return pd.to_datetime(np.random.randint(start_u, end_u, n), unit='s').floor('d')


d1 = random_dates(pd.to_datetime('2015-01-01'), pd.to_datetime('2018-01-01')) + pd.offsets.MonthBegin(0)
d2 = random_dates(pd.to_datetime('2018-01-01'), pd.to_datetime('2020-01-01')) + pd.offsets.MonthBegin(0)

df = pd.DataFrame({'lower_month':d1, 'upper_month':d2})
rng = pd.date_range('2015-01-01', freq='MS', periods=6 * 12)
vals = list(zip(df['lower_month'], df['upper_month']))

In [238]: %timeit pd.Series({y: [y >= x1 and y <= x2 for x1, x2 in vals].count(True) for y in rng})
158 ms ± 2.55 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

In [239]: %timeit pd.Series({y: sum(y >= x1 and y <= x2 for x1, x2 in vals) for y in rng})
221 ms ± 17 ms per loop (mean ± std. dev. of 7 runs, 10 loops each)

#first solution is slow    
In [240]: %timeit pd.DataFrame([(y, y >= x1 and y <= x2) for x1, x2 in vals for y in rng],                  columns=['d','test']).groupby('d')['test'].sum().astype(int)
4.52 s ± 396 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)

Answer 2

I have used itertools to repeat the upper_month and lower month for each index_date

Then compare the index_date for each lower and upper_month and set temporary column

check = 1

Then sum check after group by index_date

import pandas as pd
from pandas.compat import StringIO, BytesIO
import itertools

#sample data
data = ('contributor_id,timestamp,edits,upper_month,lower_month\n'
'8,2018-01-01,1,2018-04-01,2018-02-01\n'
'26424341,2018-01-01,11,2018-04-01,2018-02-01\n'
'26870381,2018-02-01,465,2018-04-01,2018-02-01\n'
'28109145,2018-03-01,17,2018-06-01,2018-04-01\n')

orig_df = pd.read_csv(StringIO(data))

# sample index_dates
index_df = list(pd.Series(["2018-01-01", "2018-02-01"]))

# repeat upper_month and lower_month using itertools.product
abc = list(orig_df[['upper_month','lower_month']].values)
combine_list = [index_df,abc]
res = list(itertools.product(*combine_list))
df = pd.DataFrame(res,columns=["timestamp","range"])

#separate lower_month and upper_month from  range 
df['lower_month'] = df['range'].apply(lambda x : x[1])
df['upper_month'] = df['range'].apply(lambda x : x[0])
df.drop(['range'],axis=1,inplace=True)

# convert all dates column to make them consistent
orig_df['timestamp'] = pd.to_datetime(orig_df['timestamp']).dt.date.astype(str)
orig_df['upper_month'] = pd.to_datetime(orig_df['upper_month']).dt.date.astype(str)
orig_df['lower_month'] = pd.to_datetime(orig_df['lower_month']).dt.date.astype(str)

#apply condition to set check 1
df.loc[(df["timestamp"]>=df['lower_month']) & (df["timestamp"]<=df['upper_month']),"check"] = 1

#simply groupby to count check
res = df.groupby(['timestamp'])['check'].sum()

print(res)

timestamp
2018-01-01    0.0
2018-02-01    3.0