Cleaning `Inf` values from an R dataframe

In R, I have an operation which creates some Inf values when I transform a dataframe.

I would like to turn these Inf values into NA values. The code I have is slow for large data, is there a faster way of doing this?

Say I have the following dataframe:

dat <- data.frame(a=c(1, Inf), b=c(Inf, 3), d=c("a","b"))

The following works in a single case:

 dat[,1][is.infinite(dat[,1])] = NA

So I generalized it with following loop

cf_DFinf2NA <- function(x)
{
    for (i in 1:ncol(x)){
          x[,i][is.infinite(x[,i])] = NA
    }
    return(x)
}

But I don't think that this is really using the power of R.

Solution 1:

Option 1

Use the fact that a data.frame is a list of columns, then use do.call to recreate a data.frame.

do.call(data.frame,lapply(DT, function(x) replace(x, is.infinite(x),NA)))

Option 2 -- data.table

You could use data.table and set. This avoids some internal copying.

DT <- data.table(dat)
invisible(lapply(names(DT),function(.name) set(DT, which(is.infinite(DT[[.name]])), j = .name,value =NA)))

Or using column numbers (possibly faster if there are a lot of columns):

for (j in 1:ncol(DT)) set(DT, which(is.infinite(DT[[j]])), j, NA)

Timings

# some `big(ish)` data
dat <- data.frame(a = rep(c(1,Inf), 1e6), b = rep(c(Inf,2), 1e6), 
                  c = rep(c('a','b'),1e6),d = rep(c(1,Inf), 1e6),  
                  e = rep(c(Inf,2), 1e6))
# create data.table
library(data.table)
DT <- data.table(dat)

# replace (@mnel)
system.time(na_dat <- do.call(data.frame,lapply(dat, function(x) replace(x, is.infinite(x),NA))))
## user  system elapsed 
#  0.52    0.01    0.53 

# is.na (@dwin)
system.time(is.na(dat) <- sapply(dat, is.infinite))
# user  system elapsed 
# 32.96    0.07   33.12 

# modified is.na
system.time(is.na(dat) <- do.call(cbind,lapply(dat, is.infinite)))
#  user  system elapsed 
# 1.22    0.38    1.60 


# data.table (@mnel)
system.time(invisible(lapply(names(DT),function(.name) set(DT, which(is.infinite(DT[[.name]])), j = .name,value =NA))))
# user  system elapsed 
# 0.29    0.02    0.31 

data.table is the quickest. Using sapply slows things down noticeably.

Solution 2:

Use sapply and is.na<-

> dat <- data.frame(a=c(1, Inf), b=c(Inf, 3), d=c("a","b"))
> is.na(dat) <- sapply(dat, is.infinite)
> dat
   a  b d
1  1 NA a
2 NA  3 b

Or you can use (giving credit to @mnel, whose edit this is),

> is.na(dat) <- do.call(cbind,lapply(dat, is.infinite))

which is significantly faster.

Solution 3:

Here is a dplyr/tidyverse solution using the na_if() function:

dat %>% mutate_if(is.numeric, list(~na_if(., Inf)))

Note that this only replaces positive infinity with NA. Need to repeat if negative infinity values also need to be replaced.

dat %>% mutate_if(is.numeric, list(~na_if(., Inf))) %>% 
  mutate_if(is.numeric, list(~na_if(., -Inf)))

Solution 4:

[<- with mapply is a bit faster than sapply.

> dat[mapply(is.infinite, dat)] <- NA

With mnel's data, the timing is

> system.time(dat[mapply(is.infinite, dat)] <- NA)
#   user  system elapsed 
# 15.281   0.000  13.750 

Solution 5:

There is very simple solution to this problem in the hablar package:

library(hablar)

dat %>% rationalize()

Which return a data frame with all Inf are converted to NA.

Timings compared to some above solutions. Code: library(hablar) library(data.table)

dat <- data.frame(a = rep(c(1,Inf), 1e6), b = rep(c(Inf,2), 1e6), 
                  c = rep(c('a','b'),1e6),d = rep(c(1,Inf), 1e6),  
                  e = rep(c(Inf,2), 1e6))
DT <- data.table(dat)

system.time(dat[mapply(is.infinite, dat)] <- NA)
system.time(dat[dat==Inf] <- NA)
system.time(invisible(lapply(names(DT),function(.name) set(DT, which(is.infinite(DT[[.name]])), j = .name,value =NA))))
system.time(rationalize(dat))

Result:

> system.time(dat[mapply(is.infinite, dat)] <- NA)
   user  system elapsed 
  0.125   0.039   0.164 
> system.time(dat[dat==Inf] <- NA)
   user  system elapsed 
  0.095   0.010   0.108 
> system.time(invisible(lapply(names(DT),function(.name) set(DT, which(is.infinite(DT[[.name]])), j = .name,value =NA))))
   user  system elapsed 
  0.065   0.002   0.067 
> system.time(rationalize(dat))
   user  system elapsed 
  0.058   0.014   0.072 
> 

Seems like data.table is faster than hablar. But has longer syntax.