I have a long dataframe with an index of a timeseries like this:
datetime number
2015-07-06 00:00:00 12
2015-07-06 00:10:00 55
2015-07-06 00:20:00 129
2015-07-06 00:30:00 5
2015-07-06 00:40:00 3017
2015-07-06 00:50:00 150
2015-07-06 01:00:00 347
2015-07-06 01:10:00 8
2015-07-06 01:20:00 19
... ...
I would like to transform/reshape this by splitting the column every n rows into a row in a 'new' table.
For example, an n=3 create:
datetime #0 #1 #2
2015-07-06 00:00:00 12 55 129
2015-07-06 00:30:00 5 3017 150
2015-07-06 01:00:00 347 8 19
... ... ... ...
I can think of doing this with a For-Loop, but I was wondering if there was a more efficient way native to Pandas.
You can use groupby and apply/agg with list:
u = df.groupby(pd.Grouper(key='datetime', freq='30min'))['number'].agg(list)
pd.DataFrame(u.tolist(), index=u.index)
0 1 2
datetime
2015-07-06 00:00:00 12 55 129
2015-07-06 00:30:00 5 3017 150
2015-07-06 01:00:00 347 8 19
Here is one solution
n = 3
new_df = df.groupby(df.index//n).agg({'datetime': 'first', 'number': lambda x: x.tolist()})
new_df.assign(**(new_df.number.apply(pd.Series).add_prefix('#')))
datetime number #0 #1 #2
0 2015-07-06 00:00:00 [12, 55, 129] 12 55 129
1 2015-07-06 00:30:00 [5, 3017, 150] 5 3017 150
2 2015-07-06 01:00:00 [347, 8, 19] 347 8 19
You can drop the number column
Edit: As #coldspeed suggested, you can combine the last two steps.
new_df = df.groupby(df.index//n).agg({'datetime': 'first', 'number': lambda x: x.tolist()})
new_df.assign(**(new_df.pop('number').apply(pd.Series).add_prefix('#')))
datetime #0 #1 #2
0 2015-07-06 00:00:00 12 55 129
1 2015-07-06 00:30:00 5 3017 150
2 2015-07-06 01:00:00 347 8 19
Related
I have a dataframe that consists of 3 years of data and two columns remaining useful life and predicted remaining useful life.
I am aggregating rul and pred_rul of 3 years data for each machineID for the maximum date they have. The original dataframe looks like this-
rul pred_diff machineID datetime
10476749 870 312.207825 408 2021-05-25 00:00:00
11452943 68 288.517578 447 2023-03-01 12:00:00
12693829 381 273.159698 493 2021-09-16 16:00:00
3413787 331 291.326416 133 2022-10-26 12:00:00
464093 77 341.506195 19 2023-10-10 16:00:00
... ... ... ... ...
11677555 537 310.586090 456 2022-04-07 00:00:00
2334804 551 289.307129 92 2021-09-04 20:00:00
5508311 35 293.721771 214 2023-01-06 04:00:00
12319704 348 322.199219 479 2021-11-11 20:00:00
4777501 87 278.089417 186 2021-06-29 12:00:00
1287421 rows × 4 columns
And I am aggregating it based on this code-
y_test_grp = y_test.groupby('machineID').agg({'datetime':'max', 'rul':'mean', 'pred_diff':'mean'})[['datetime','rul', 'pred_diff']].reset_index()
which gives the following output-
machineID datetime rul pred_diff
0 1 2023-10-03 20:00:00 286.817681 266.419401
1 2 2023-11-14 00:00:00 225.561953 263.372531
2 3 2023-10-25 00:00:00 304.736237 256.933351
3 4 2023-01-13 12:00:00 204.084899 252.476066
4 5 2023-09-07 00:00:00 208.702431 252.487156
... ... ... ... ...
495 496 2023-10-11 00:00:00 302.445285 298.836798
496 497 2023-08-26 04:00:00 281.601613 263.479885
497 498 2023-11-28 04:00:00 292.593906 263.985034
498 499 2023-06-29 20:00:00 260.887529 263.494844
499 500 2023-11-08 20:00:00 160.223614 257.326034
500 rows × 4 columns
Since this is grouped by on machineID, it is giving just 500 rows which is less. I want to aggregate rul and pred_rul on weekly basis such that for each machineID I get 52weeks*3years=156 rows. I am not able to identify which function to use for taking 7 days as interval and aggregating rul and pred_rul on that.
You can use Grouper:
pd.groupby(['machineID', pd.Grouper(key='datetime', freq='7D')]).mean()
I have these data:
val1 val2 val3
dt
2017-12-15 00:00:00 81 90 79
2017-12-15 00:01:00 67 85 80
2017-12-15 00:02:00 4 41 37
2017-12-15 00:03:00 61 68 29
2017-12-15 00:04:00 49 6 56
2017-12-15 00:05:00 94 13 93
2017-12-15 00:06:00 91 3 75
2017-12-15 00:07:00 94 81 7
2017-12-15 00:08:00 55 59 33
2017-12-15 00:09:00 97 89 26
2017-12-15 00:10:00 17 75 88
2017-12-15 00:11:00 39 40 96
2017-12-15 00:12:00 61 20 70
2017-12-15 00:13:00 62 31 93
2017-12-15 00:14:00 7 26 29
I would like to find the 3 max values for each 5-minute period.
The max values can be in any column (val1, val2, val3) and must be searched among the 15 values available for the 5 minutes.
At the moment I can only find the largest in a single column.
Is it possible to search for nlargest in multiple columns?
This is the code to generate the data and to search for the max for val1:
import pandas as pd
import numpy as np
from datetime import datetime, timedelta
date_ref = datetime(2017, 12, 15, 0,0,0)
dtime = pd.date_range(date_ref, freq='1min', periods=15)
np.random.seed(seed=1115)
data1 = np.random.randint(1, high=100, size=len(dtime))
data2 = np.random.randint(1, high=100, size=len(dtime))
data3 = np.random.randint(1, high=100, size=len(dtime))
df = pd.DataFrame({'dt': dtime, 'val1': data1, 'val2': data2, 'val3': data3})
df.set_index('dt', inplace=True)
print(df)
group = df.groupby(pd.Grouper(freq='5min'))
max_only_for_val1 = (pd.DataFrame(
group["val1"]
.nlargest(3))
.reset_index(level=1, drop=True)
)
print(max_only_for_val1)
This is the output:
val1
dt
2017-12-15 00:00:00 81
2017-12-15 00:00:00 67
2017-12-15 00:00:00 61
2017-12-15 00:05:00 97
2017-12-15 00:05:00 94
2017-12-15 00:05:00 94
2017-12-15 00:10:00 62
2017-12-15 00:10:00 61
2017-12-15 00:10:00 39
Since it doesn't matter where your values come from, let's reshape your data a bit.
df = df.reset_index().melt('dt').drop('variable', 1)
df.head(10)
dt value
0 2017-12-15 00:00:00 81
1 2017-12-15 00:01:00 67
2 2017-12-15 00:02:00 4
3 2017-12-15 00:03:00 61
4 2017-12-15 00:04:00 49
5 2017-12-15 00:05:00 94
6 2017-12-15 00:06:00 91
7 2017-12-15 00:07:00 94
8 2017-12-15 00:08:00 55
9 2017-12-15 00:09:00 97
Now, call groupby + apply -
def get_max3(x):
return x.sort_values(ascending=False).head(3)
df = df.groupby(pd.Grouper(key='dt', freq='5min'))['value']\
.apply(get_max3)\
.reset_index(0)\
.reset_index(drop=True)
dt value
0 2017-12-15 00:00:00 90
1 2017-12-15 00:00:00 85
2 2017-12-15 00:00:00 81
3 2017-12-15 00:05:00 97
4 2017-12-15 00:05:00 94
5 2017-12-15 00:05:00 94
6 2017-12-15 00:10:00 96
7 2017-12-15 00:10:00 93
8 2017-12-15 00:10:00 88
An alternative definition for get_max3 using numpy.sort -
def get_max3(x):
return np.sort(x.values)[-4::-1]
A B C D yearweek
0 245 95 60 30 2014-48
1 245 15 70 25 2014-49
2 150 275 385 175 2014-50
3 100 260 170 335 2014-51
4 580 925 535 2590 2015-02
5 630 126 485 2115 2015-03
6 425 90 905 1085 2015-04
7 210 670 655 945 2015-05
The last column contains the the year along with the weeknumber. Is it possible to convert this to a datetime column with pd.to_datetime?
I've tried:
pd.to_datetime(df.yearweek, format='%Y-%U')
0 2014-01-01
1 2014-01-01
2 2014-01-01
3 2014-01-01
4 2015-01-01
5 2015-01-01
6 2015-01-01
7 2015-01-01
Name: yearweek, dtype: datetime64[ns]
But that output is incorrect, while I believe %U should be the format string for week number. What am I missing here?
You need another parameter for specify day - check this:
df = pd.to_datetime(df.yearweek.add('-0'), format='%Y-%W-%w')
print (df)
0 2014-12-07
1 2014-12-14
2 2014-12-21
3 2014-12-28
4 2015-01-18
5 2015-01-25
6 2015-02-01
7 2015-02-08
Name: yearweek, dtype: datetime64[ns]
I have a fairly complicated question. I need to select rows from a data frame within a certain set of start and end dates, and then sum those values and put them in a new dataframe.
So I start off with with data frame, df:
import random
dates = pd.date_range('20150101 020000',periods=1000)
df = pd.DataFrame({'_id': random.choice(range(0, 1000)),
'time_stamp': dates,
'value': random.choice(range(2,60))
})
and define some start and end dates:
import pandas as pd
start_date = ["2-13-16", "2-23-16", "3-17-16", "3-24-16", "3-26-16", "5-17-16", "5-25-16", "10-10-16", "10-18-16", "10-23-16", "10-31-16", "11-7-16", "11-14-16", "11-22-16", "1-23-17", "1-29-17", "2-06-17", "3-11-17", "3-23-17", "6-21-17", "6-28-17"]
end_date = pd.DatetimeIndex(start_date) + pd.DateOffset(7)
Then what needs to happen is that I need to create a new data frame with weekly_sum which sums the value column of df which occur in between the the start_date and end_date.
So for example, the first row of the new data frame would return the sum of the values between 2-13-16 and 2-20-16. I imagine I'd use groupby.sum() or something similar.
It might look like this:
id start_date end_date weekly_sum
65 2016-02-13 2016-02-20 100
Any direction is greatly appreciated!
P.S. I know my use of random.choice is a little wonky so if you have a better way of generating random numbers, I'd love to see it!
You can use
def get_dates(x):
# Select the df values between start and ending datetime.
n = df[(df['time_stamp']>x['start'])&(df['time_stamp']<x['end'])]
# Return first id and sum of values
return n['id'].values[0],n['value'].sum()
dates = pd.date_range('20150101 020000',periods=1000)
df = pd.DataFrame({'id': np.random.randint(0,1000,size=(1000,)),
'time_stamp': dates,
'value': np.random.randint(2,60,size=(1000,))
})
ndf = pd.DataFrame({'start':pd.to_datetime(start_date),'end':end_date})
#Unpack and assign values to id and value column
ndf[['id','value']] = ndf.apply(lambda x : get_dates(x),1).apply(pd.Series)
print(df.head(5))
id time_stamp value
0 770 2015-01-01 02:00:00 59
1 781 2015-01-02 02:00:00 32
2 761 2015-01-03 02:00:00 40
3 317 2015-01-04 02:00:00 16
4 538 2015-01-05 02:00:00 20
print(ndf.head(5))
end start id value
0 2016-02-20 2016-02-13 569 221
1 2016-03-01 2016-02-23 28 216
2 2016-03-24 2016-03-17 152 258
3 2016-03-31 2016-03-24 892 265
4 2016-04-02 2016-03-26 606 244
You can calculate a weekly summary with the following code. The code below is based on Monday.
import pandas as pd
import random
dates = pd.date_range('20150101 020000',periods=1000)
df = pd.DataFrame({'_id': random.choice(range(0, 1000)),
'time_stamp': dates,
'value': random.choice(range(2,60))
})
df['day_of_week'] = df['time_stamp'].dt.weekday_name
df['start'] = np.where(df["day_of_week"]=="Monday", 1, 0)
df['week'] = df["start"].cumsum()
# It is based on Monday.
df.head(20)
# Out[109]:
# _id time_stamp value day_of_week start week
# 0 396 2015-01-01 02:00:00 59 Thursday 0 0
# 1 396 2015-01-02 02:00:00 59 Friday 0 0
# 2 396 2015-01-03 02:00:00 59 Saturday 0 0
# 3 396 2015-01-04 02:00:00 59 Sunday 0 0
# 4 396 2015-01-05 02:00:00 59 Monday 1 1
# 5 396 2015-01-06 02:00:00 59 Tuesday 0 1
# 6 396 2015-01-07 02:00:00 59 Wednesday 0 1
# 7 396 2015-01-08 02:00:00 59 Thursday 0 1
# 8 396 2015-01-09 02:00:00 59 Friday 0 1
# 9 396 2015-01-10 02:00:00 59 Saturday 0 1
# 10 396 2015-01-11 02:00:00 59 Sunday 0 1
# 11 396 2015-01-12 02:00:00 59 Monday 1 2
# 12 396 2015-01-13 02:00:00 59 Tuesday 0 2
# 13 396 2015-01-14 02:00:00 59 Wednesday 0 2
# 14 396 2015-01-15 02:00:00 59 Thursday 0 2
# 15 396 2015-01-16 02:00:00 59 Friday 0 2
# 16 396 2015-01-17 02:00:00 59 Saturday 0 2
# 17 396 2015-01-18 02:00:00 59 Sunday 0 2
# 18 396 2015-01-19 02:00:00 59 Monday 1 3
# 19 396 2015-01-20 02:00:00 59 Tuesday 0 3
aggfunc = {"time_stamp": [np.min, np.max], "value": [np.sum]}
df2 = df.groupby("week", as_index=False).agg(aggfunc)
df2.columns = ["week", "start_date", "end_date", "weekly_sum"]
df2.iloc[58:61]
# Out[110]:
# week start_date end_date weekly_sum
# 58 58 2016-02-08 02:00:00 2016-02-14 02:00:00 413
# 59 59 2016-02-15 02:00:00 2016-02-21 02:00:00 413
# 60 60 2016-02-22 02:00:00 2016-02-28 02:00:00 413
I have the table below in a Pandas dataframe:
date user_id whole_cost cost1
02/10/2012 00:00:00 1 1790 12
07/10/2012 00:00:00 1 364 15
30/01/2013 00:00:00 1 280 10
02/02/2013 00:00:00 1 259 24
05/03/2013 00:00:00 1 201 39
02/10/2012 00:00:00 3 623 1
07/12/2012 00:00:00 3 90 0
30/01/2013 00:00:00 3 312 90
02/02/2013 00:00:00 5 359 45
05/03/2013 00:00:00 5 301 34
02/02/2013 00:00:00 5 359 1
05/03/2013 00:00:00 5 801 12
..
The table was extracted from a csv file using the following query :
import pandas as pd
newnames = ['date','user_id', 'whole_cost', 'cost1']
df = pd.read_csv('expenses.csv', names = newnames, index_col = 'date')
I have to analyse the profile of my users and for this purpose:
I would like to group (for each user - they are thousands) queries by month summing the query whole_cost for the entire month e.g. if user_id=1 was has a whole cost of 1790 on 02/10/2012 with cost1 12 and on the 07/10/2012 with whole cost 364, then it should have an entry in the new table of 2154 (as the whole cost) on 31/10/2012 (end of the month end-point representing the month - all dates in the transformed table will be month ends representing the whole month to which they relate).
In 0.14 you'll be able to groupby monthly and another column at the same time:
In [11]: df
Out[11]:
user_id whole_cost cost1
2012-10-02 1 1790 12
2012-10-07 1 364 15
2013-01-30 1 280 10
2013-02-02 1 259 24
2013-03-05 1 201 39
2012-10-02 3 623 1
2012-12-07 3 90 0
2013-01-30 3 312 90
2013-02-02 5 359 45
2013-03-05 5 301 34
2013-02-02 5 359 1
2013-03-05 5 801 12
In [12]: df1 = df.sort_index() # requires sorted DatetimeIndex
In [13]: df1.groupby([pd.TimeGrouper(freq='M'), 'user_id'])['whole_cost'].sum()
Out[13]:
user_id
2012-10-31 1 2154
3 623
2012-12-31 3 90
2013-01-31 1 280
3 312
2013-02-28 1 259
5 718
2013-03-31 1 201
5 1102
Name: whole_cost, dtype: int64
until 0.14 I think you're stuck with doing two groupbys:
In [14]: g = df.groupby('user_id')['whole_cost']
In [15]: g.resample('M', how='sum').dropna()
Out[15]:
user_id
1 2012-10-31 2154
2013-01-31 280
2013-02-28 259
2013-03-31 201
3 2012-10-31 623
2012-12-31 90
2013-01-31 312
5 2013-02-28 718
2013-03-31 1102
dtype: float64
With timegrouper getting deprecated, you can replace it with Grouper to get the same results
df.groupby(['user_id', pd.Grouper(key='date', freq='M')]).agg({'whole_cost':sum})
df.groupby(['user_id', df['date'].dt.dayofweek]).agg({'whole_cost':sum})