Repeatedly apply a conditional summary to groups in a dataframe - r

Question

I have a large dataframe that looks like this:
group_id distance metric
1 1.1 0.85
1 1.1 0.37
1 1.7 0.93
1 2.3 0.45
...
1 6.3 0.29
1 7.9 0.12
2 2.5 0.78
2 2.8 0.32
...
The dataframe is already sorted by group_id and then distance. I want know the dplyr or data.table efficient equivalent to doing the following operations:
Within each group_id:
Let the unique and sorted values of distance within the current group_id be d1,d2,...,d_n.
For each d in d1,d2,...,d_n: Compute some function f on all values of metric whose distance value is less than d. The function f is a custom user defined function, that takes in a vector and returns a scalar. Assume that the function f is well defined on an empty vector.
So, in the example above, the desired dataframe would look like:
group_id distance_less_than metric
1 1.1 f(empty vector)
1 1.7 f(0.85, 0.37)
1 2.3 f(0.85, 0.37, 0.93)
...
1 7.9 f(0.85, 0.37, 0.93, 0.45,...,0.29)
2 2.5 f(empty vector)
2 2.8 f(0.78)
...
Notice how distance values can be repeated, like the value 1.1 under group 1. In such cases, both of the rows should be excluded when the distance is less than 1.1 (in this case this results in an empty vector).

Answer 1

A possible approach is to use non-equi join available in data.table. The left table is the unique set of combinations of group_id and distance and right table are all the distance less than left table's distance.
f <- sum
DT[unique(DT, by=c("group_id", "distance")), on=.(group_id, distance<distance), allow.cartesian=TRUE,
f(metric), by=.EACHI]
output:
group_id distance V1
1: 1 1.1 NA
2: 1 1.7 1.22
3: 1 2.3 2.15
4: 1 6.3 2.60
5: 1 7.9 2.89
6: 2 2.5 NA
7: 2 2.8 0.78
data:
library(data.table)
DT <- fread("group_id distance metric
1 1.1 0.85
1 1.1 0.37
1 1.7 0.93
1 2.3 0.45
1 6.3 0.29
1 7.9 0.12
2 2.5 0.78
2 2.8 0.32")

Answer 2

Don't think this would be faster than data.table option but here is one way using dplyr
library(dplyr)
df %>%
group_by(group_id) %>%
mutate(new = purrr::map_dbl(distance, ~f(metric[distance < .])))
where f is your function. map_dbl expects return type of function to be double. If you have different return type for your function you might want to use map_int, map_chr or likes.
If you want to keep only one entry per distance you might remove them using filter and duplicated
df %>%
group_by(group_id) %>%
mutate(new = purrr::map_dbl(distance, ~f(metric[distance < .]))) %>%
filter(!duplicated(distance))