bench pour compter le nombre de NA en lignes

bench.R

library(purrr)
library(charlatan)
library(microbenchmark)


create_fake_data <- function(n){
  data <- ch_generate('job', 'color_name', 'phone_number', n = n)
  
  data$color_name[sample.int(nrow(data), n/10 )] <- NA_character_
  data$x <- rnorm(n)
  data$x[sample.int(nrow(data), n/10 )] <- NA_real_
  
  data$y <- rpois(n = n, lambda = 12)
  data$y[sample.int(nrow(data), n/10 )] <- NA_real_
  data  
}


prop0 <- function(dataset){
  rowSums(is.na(dataset))
}

prop1 <- function(dataset){
  # probleme de cast ; quand les données contiennent des colonnes de types differents 
  # il y aura changement de type - les NA seront castes correctement en NA mais 
  # surement que cela fait une operation non necessaire
  apply(dataset, MARGIN = 1, function(x){
    sum(is.na(x))
  })
}

prop2 <- function(dataset){
  # probleme, l'allocation d'un nouveau vecteur 
  # pour chaque ligne va deteriorer les performances + cast
  # quand colonnes de differents types + purrr fait quelques 
  # operations de plus que sapply 
  pmap_dbl(dataset, function(...){sum(is.na(c(...)))})
}

prop3 <- function(dataset){
  # probleme de cette solution, l'allocation d'un nouveau vecteur 
  # pour chaque ligne va deteriorer les performances + cast
  # quand colonnes de differents types
  dataset <- as.list(dataset)
  dataset$FUN <- function(...){sum(is.na(c(...)))}
  dataset$SIMPLIFY <- TRUE
  dataset$USE.NAMES = FALSE
  do.call(mapply, dataset)
}

prop4 <- function(dataset){
  # moi j'aime bien celle la
  Reduce(`+`, lapply(dataset, is.na) )
}

prop5 <- function(df) {
  # celle de Florian - proche de la precedente
  nb_na <- integer(nrow(df))
  for (var in df) {
    nb_na <- nb_na + is.na(var)
  }
  nb_na
}

bench_that <- function(x){
  
  message("Le dataset contient ", nrow(x), " lignes")
  
  bench <- microbenchmark(times = 50L, 
                          `Reduce sur colonnes` = prop4(x),
                          `rowSums(...)` = prop0(x), 
                          `apply(..., MARGIN = 1)` = prop1(x), 
                          `sapply on one vector per row` = prop3(x), 
                          `purrr on one vector per row` = prop2(x), 
                          `florian` = prop5(x), 
                          `naniar:::n_miss_row(data1)` = naniar:::n_miss_row(x)
  )
  print(bench)
  bench
}

data1 <- create_fake_data(1000)
data2 <- create_fake_data(10000)
data3 <- create_fake_data(100000)

b1 <- bench_that(data1)
b2 <- bench_that(data2)
b3 <- bench_that(data3)

benchnames <- c("b1", "b2", "b3")
benchbars <- c("bar1", "bar2", "bar3")

bench <- Reduce(rbind, lapply(benchnames, 
                              function(x) {
                                obj <- get(x)
                                data.frame(expr = obj$expr, time = obj$time/1000, bench = x)
                              } ) )

tab <- with(bench, tapply(time, list(expr = expr, bench = bench), median, na.rm = TRUE))
print(tab)

                              bench
expr                                  b1         b2         b3
  Reduce sur colonnes            36.6345   195.8080   3093.197
  rowSums(...)                   57.1660   249.5275   3650.098
  apply(..., MARGIN = 1)       2731.2415 25391.1510 380552.698
  sapply on one vector per row 5247.7820 63583.2755 745344.002
  purrr on one vector per row  5767.7365 66034.7290 763841.574
  florian                        34.6465   209.8205   3675.005
  naniar:::n_miss_row(data1)     71.2370   272.3780   3956.702

Reduce(+, lapply(dataset, is.na) ) is the faster way to count NA per rows. Florian for loop is almost the same.

If using prop5, you won't face the following issue:

> Reduce(`+`, lapply(iris[0], is.na) )
NULL

> prop5(iris[0])
  [1] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
 [57] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0
[113] 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0 0