CODEWITHSUNDEEP
rdplyrdata.tablepurrr
Geet
Geet

Reputation: 2575

Calculate all variable combinations and the distinct counts of df using R data.table

How can I run comboGeneral with all possible m using data.table to get all possible variable combinations? Then, how can I calculate the distinct count in all the dataframes subsetted using these variable combinations?

Here is a purrr and dplyr version. I need nms and counts using data.table.

library(data.table); library(dplyr); library(magrittr); library(RcppAlgos); library(purrr)

num_m <- seq_len(ncol(mtcars))
nam_list <- names(mtcars)

nms <- map(num_m, ~comboGeneral(nam_list, m = .x, FUN = c)) %>% unlist(recursive = FALSE)

counts <- map_dbl(nms, ~(mtcars %>% select(.x) %>% n_distinct()))

Upvotes: 2

Views: 165

Answers (1)

MichaelChirico
MichaelChirico

Reputation: 34763

It's not clear what you're hoping to accomplish by using data.table specifically for the first part. comboGeneral is from RccpAlgos so I assume it's optimized pretty heavily... combn in base R is the alternative (this is not really something data.table would have any implementation for...):

nms = unlist(lapply(num_m, combn, x = nam_list, simplify = FALSE), recursive = FALSE)

With that in hand, there's a few ways in data.table:

mtcars = as.data.table(mtcars)
counts = sapply(nms, uniqueN, x = mtcars)

Or

sapply(nms, function(nm) nrow(mtcars[ , TRUE, keyby = nm]))

Or

sapply(nms, function(nm) nrow(unique(mtcars, by = nm)))

It seems that the first option is not only the most concise but also the most efficient:

library(microbenchmark)

microbenchmark(times = 100L,
               map_dbl(nms, ~(mtcars %>% select(.x) %>% n_distinct())),
               sapply(nms, uniqueN, x = mtcars),
               sapply(nms, function(nm) nrow(mtcars[ , TRUE, keyby = nm])),
               sapply(nms, function(nm) nrow(unique(mtcars, by = nm))))
# Unit: milliseconds
#                                                        expr        min         lq
#     map_dbl(nms, ~(mtcars %>% select(.x) %>% n_distinct())) 2246.10862 2365.33801
#                            sapply(nms, uniqueN, x = mtcars)   66.16144   68.95391
#  sapply(nms, function(nm) nrow(mtcars[, TRUE, keyby = nm])) 1659.20425 1701.79188
#     sapply(nms, function(nm) nrow(unique(mtcars, by = nm)))  102.42203  106.87100
#        mean     median         uq       max neval
#  2469.50648 2448.44821 2544.00350 3530.6513   100
#    73.28518   71.54861   75.85161  118.5919   100
#  1796.30372 1766.59618 1825.97374 2881.2376   100
#   113.63032  111.28377  118.22441  174.2691   100

Regarding speeding up the first step, you can get about 10% speed-up by dropping the sugar of map and going for raw lapply:

microbenchmark(times = 1000L,
               lapply(num_m, combn, x = nam_list, simplify = FALSE),
               map(num_m, ~comboGeneral(nam_list, m = .x, FUN = c)),
               lapply(num_m, function(m) comboGeneral(nam_list, m, FUN = c)))
# Unit: microseconds
#                                                           expr      min        lq
#           lapply(num_m, combn, x = nam_list, simplify = FALSE) 1718.994 1847.3710
#           map(num_m, ~comboGeneral(nam_list, m = .x, FUN = c))  564.076  629.5120
#  lapply(num_m, function(m) comboGeneral(nam_list, m, FUN = c))  473.135  525.2655
#       mean    median        uq      max neval
#  2088.7454 1921.8840 2016.0275 7789.501  1000
#   713.8342  661.0455  709.4650 3800.253  1000
#   593.7732  550.2460  583.7005 5190.982  1000

Note: We cannot use lapply(num_m, comboGeneral, v = nam_list, FUN = c) because FUN will be interpreted as the argument to lapply, not to comboGeneral.

Upvotes: 3

Related Questions