Introduction

This time, the word "event data" is used to mean data such as ** (time, event name (, user) (, quantity)) **. When trying to create a user prediction model, you may want to format such event data so that it is "one user, one line".

Make a note of how pandas handles this area.

Method

First, consider the following event data as an example.

import pandas as pd

data = [
    ['user1', '2015-7-1 10:00:00', 'login'],
    ['user1', '2015-7-1 10:00:10', 'item'],
    ['user1', '2015-7-1 10:00:30', 'item'],
    ['user2', '2015-7-1 10:01:00', 'top'],
    ['user1', '2015-7-1 10:01:30', 'cart'],
    ['user2', '2015-7-1 10:01:50', 'login'],
    ['user2', '2015-7-1 10:02:30', 'logout'],

    ['user1', '2015-7-2 13:00:00', 'login'],
    ['user1', '2015-7-2 13:01:00', 'logout'],
    ['user3', '2015-7-2 13:01:00', 'top'],
    ['user2', '2015-7-2 13:01:50', 'login'],
    ['user2', '2015-7-2 13:02:30', 'item'],
    ['user2', '2015-7-2 13:03:30', 'cart'],
    ['user2', '2015-7-2 13:03:30', 'history'],
]
df = pd.DataFrame(data)
df.columns = ['user_id', 'time', 'event']
df['time'] = pd.to_datetime(df['time'])
df['dummy'] = 1
df

	user_id	time	event	dummy
0	user1	2015-07-01 10:00:00	login	1
1	user1	2015-07-01 10:00:10	item	1
2	user1	2015-07-01 10:00:30	item	1
3	user2	2015-07-01 10:01:00	top	1
4	user1	2015-07-01 10:01:30	cart	1
5	user2	2015-07-01 10:01:50	login	1
6	user2	2015-07-01 10:02:30	logout	1
7	user1	2015-07-02 13:00:00	login	1
8	user1	2015-07-02 13:01:00	logout	1
9	user3	2015-07-02 13:01:00	top	1
10	user2	2015-07-02 13:01:50	login	1
11	user2	2015-07-02 13:02:30	item	1
12	user2	2015-07-02 13:03:30	cart	1
13	user2	2015-07-02 13:03:30	history	1

Suppose you want to format this data as "1 user, 1 row" and "the total number of occurrences for each daily event" as a column.

pivot_table and Grouper

You can use the DataFrame class pivot_table () to do something similar to an Excel PivotTable. I'd like to aggregate by day, but if I use Grouper (), I can group the datetime type columns nicely. How convenient ...

res2 = df.pivot_table(index=['user_id', pd.Grouper(freq='d', key='time')], columns='event', values='dummy', aggfunc=len)
res2

	event	cart	history	item	login	logout	top
user_id	time
user1	2015-07-01	1	NaN	2	1	NaN	NaN
user1	2015-07-02	NaN	NaN	NaN	1	1	NaN
user2	2015-07-01	NaN	NaN	NaN	1	1	1
user2	2015-07-02	1	1	1	1	NaN	NaN
user3	2015-07-02	NaN	NaN	NaN	NaN	NaN	1

Since Grouper's freq ='d', it will be aggregated on a daily basis. If you set freq ='M', it will be monthly.

Bring the date to the column

It is good to use unstack ().

res3 = res2.unstack()
res3

event	cart		history		item		login		logout		top
time	2015-07-01	2015-07-02	2015-07-01	2015-07-02	2015-07-01	2015-07-02	2015-07-01	2015-07-02	2015-07-01	2015-07-02	2015-07-01	2015-07-02
user_id
user1	1	NaN	NaN	NaN	2	NaN	1	1	NaN	1	NaN	NaN
user2	NaN	1	NaN	1	NaN	1	1	1	1	NaN	1	NaN
user3	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	1

I want to bring the date up

Use swaplevel (Index1, which you want to replace). Also, I want to operate on the column (axis number 1), so add ʻaxis = 1`.

res4 = res3.swaplevel(0, 1, axis=1)
res4

time	2015-07-01	2015-07-02	2015-07-01	2015-07-02	2015-07-01	2015-07-02	2015-07-01	2015-07-02	2015-07-01	2015-07-02	2015-07-01	2015-07-02
event	cart	cart	history	history	item	item	login	login	logout	logout	top	top
user_id
user1	1	NaN	NaN	NaN	2	NaN	1	1	NaN	1	NaN	NaN
user2	NaN	1	NaN	1	NaN	1	1	1	1	NaN	1	NaN
user3	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	1

I want the columns to be arranged in chronological order

Use sortlevel ().

res4.sortlevel(0, axis=1)

time	2015-07-01						2015-07-02
event	cart	history	item	login	logout	top	cart	history	item	login	logout	top
user_id
user1	1	NaN	2	1	NaN	NaN	NaN	NaN	NaN	1	1	NaN
user2	NaN	NaN	NaN	1	1	1	1	1	1	1	NaN	NaN
user3	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	NaN	1

Now you have the desired shape.

at the end

There will be a column of "number of event types x number of dates", so if there are too many, it cannot be used. I think that it can be used when the number of event types is not large or when it is aggregated on a monthly basis.

Anyway, pandas is convenient.

Aggregate event data into one-user, one-line format using pandas