Aggregating by subsets in dplyr - r

Question

I have a dataset with a million records that I need to aggregate after first subsetting the data. It is difficult to provide a good reproducible sample because in this case, the sample size would be rather large - but I will try anyway.
A random sample of the data that I am working with looks like this:
> df
auto_id user_id month
164537 7124 240249 10
151635 7358 226423 9
117288 7376 172463 9
177119 6085 199194 11
128904 7110 141608 9
157194 7143 241964 9
71303 6090 141646 7
72480 6808 175910 7
108705 6602 213098 8
97889 7379 185516 8
184906 6405 212580 12
37242 6057 197905 8
157284 6548 162928 9
17910 6885 194180 10
70660 7162 161827 7
8593 7375 207061 8
28712 6311 176373 10
144194 7324 142715 9
73106 7196 176153 7
67065 7392 171039 7
77954 7116 161489 7
59842 7107 162637 7
101819 5994 182973 9
183546 6427 142029 12
102881 6477 188129 8
In every month, there many users who are the same, and first we should subset by month and make a frequency table of the users and the amount of trips taken (unfortunately, in the random sample above there is only one trip per user, but in the larger dataset, this is not the case):
full_data <- full_data[full_data$month == 7,]
users <- as.data.frame(table(full_data$user_id))
head(users)
Var1 Freq
1 100231 10
2 100744 17
3 111281 1
4 111814 2
5 113716 3
6 117493 3
As we can see, in the full data set, in month of July (month = 7), users have taken multiple trips. Now the important part - which is to subset only the top 10% of these users (the top 10% in terms of Freq)
tenPercent = round(nrow(users)/10)
users <- users[order(-users$Freq),]
topten <- head(users, n = tenPercent)
Now the new dataframe - topten - can be summed and we get the amount of trips taken by the top ten percent of users
sum(topten$Freq)
[1] 12147
In the end the output should look like this
> output
month trips
1 7 12147
2 8 ...
3 9 ...
4 10 ...
5 11 ...
6 12 ...
Is there a way to automate this process using dplyr - I mean specifically the subsetting by the top ten percent ? I have tried
output <- full_data %>%
+ group_by(month) %>%
+ summarise(n = n())
But this only aggregates total trips by month. Could someone suggest a way to integrate this part into the query in dplyr ? :
tenPercent = round(nrow(users)/10)
users <- users[order(-users$Freq),]
topten <- head(users, n = tenPercent)

Answer 1

The code below counts the number of rows for each user_id in each month, and then selects the 10% of users with the most rows in each month and sums them. Let me know if it solves your problem.
library(dplyr)
full_data %>% group_by(month, user_id) %>%
tally %>%
group_by(month) %>%
filter(percent_rank(n) >= 0.9) %>%
summarise(n_trips = sum(n))
UPDATE: Following up on your comment, let's do a check with some fake data. Below we have 30 different values of user_id and 10,000 total rows. I've also used the prob argument so that the probability of a user_id being selected is proportional to its value (i.e., user_id 1 is the least likely to be chosen and user_id 30 is the most likely to be chosen).
set.seed(3)
full_data = data.frame(user_id=sample(1:30,10000, replace=TRUE, prob=1:30),
month=sample(1:12, 10000, replace=TRUE))
Let's look as the number of rows for each user_id for month==1. The code below counts the number of rows for each user_id and sorts from most to least common. Note that the three most common values of user_id (28,29,26) comprise 171 rows (60+57+54). Since there are 30 different values of user_id the top three users represent the top 10% of users:
full_data %>% filter(month==1) %>%
group_by(month, user_id) %>%
tally %>%
arrange(desc(n)) %>% as.data.frame
month user_id n
1 1 28 60
2 1 29 57
3 1 26 54
4 1 30 53
5 1 27 49
6 1 22 43
7 1 21 41
8 1 20 40
9 1 23 40
10 1 24 38
11 1 25 38
12 1 19 37
13 1 18 33
14 1 16 28
15 1 15 27
16 1 17 27
17 1 14 26
18 1 9 20
19 1 12 20
20 1 13 20
21 1 10 17
22 1 11 17
23 1 6 15
24 1 7 13
25 1 8 13
26 1 4 9
27 1 5 7
28 1 2 3
29 1 3 2
30 1 1 1
So now let's take the next step and select the top 10% of users. To answer the question in your comment, filter(percent_rank(n) >= 0.9) keeps only the top 10% of user_id, based on the value of n (which is the number of rows for each user_id). percent_rank is on of several ranking functions in dplyr that have different ways of dealing with ties (which may be the reason you're not getting the results you expect). See ?percent_rank for details:
full_data %>% filter(month==1) %>%
group_by(month, user_id) %>%
tally %>%
group_by(month) %>%
filter(percent_rank(n) >= 0.9)
month user_id n
1 1 26 54
2 1 28 60
3 1 29 57
And the sum of n (the total number of trips for the top 10%) is:
full_data %>% filter(month==1) %>%
group_by(month, user_id) %>%
tally %>%
group_by(month) %>%
filter(percent_rank(n) >= 0.9) %>%
summarise(n_trips = sum(n))
month n_trips
1 1 171
So it looks like the code does what we'd naively expect, but maybe the issue is related to how ties are dealt with. Let me know if you're still getting anomalous results in your real data or if I've misunderstood what you're trying to accomplish.