ftransform: Fast Transform and Compute Columns on a Data Frame

Description

ftransform is a much faster version of transform for data frames. It returns the data frame with new columns computed and/or existing columns modified or deleted. settransform does all of that by reference. fcompute computes and returns new columns. These functions evaluate all arguments simultaneously, allow list-input (nested pipelines) and disregard grouped data.

Catering to the tidyverse user, v1.7.0 introduced fmutate, providing familiar functionality i.e. arguments are evaluated sequentially, computation on grouped data is done by groups, and functions can be applied to multiple columns using across. See also the Details.

Usage

# dplyr-style mutate (sequential evaluation + across() feature)
fmutate(.data, ..., .keep = "all", .cols = NULL)
mtt(.data, ..., .keep = "all", .cols = NULL) # Shorthand for fmutate
# Modify and return data frame
ftransform(.data, ...)
ftransformv(.data, vars, FUN, ..., apply = TRUE)
tfm(.data, ...)               # Shorthand for ftransform
tfmv(.data, vars, FUN, ..., apply = TRUE)
# Modify data frame by reference
settransform(.data, ...)
settransformv(.data, ...)     # Same arguments as ftransformv
settfm(.data, ...)            # Shorthand for settransform
settfmv(.data, ...)
# Replace/add modified columns in/to a data frame
ftransform(.data) <- value
tfm(.data) <- value           # Shorthand for ftransform<-
# Compute columns, returned as a new data frame
fcompute(.data, ..., keep = NULL)
fcomputev(.data, vars, FUN, ..., apply = TRUE, keep = NULL)

Value

The modified data frame .data, or, for fcompute, a new data frame with the columns computed on .data. All attributes of .data are preserved.

Arguments

.data: a data frame or named list of columns.
...: further arguments of the form column = value. The value can be a combination of other columns, a scalar value, or NULL, which deletes column. Alternatively it is also possible to place a single list here, which will be treated like a list of column = value arguments. For ftransformv and fcomputev, ... can be used to pass further arguments to FUN. The ellipsis (...) is always evaluated within the data frame (.data) environment. See Examples. fmutate additionally supports across statements, and evaluates tagged vector expressions sequentially. With grouped execution, dots can also contain arbitrary expressions that result in a list of data-length columns. See Examples.
vars: variables to be transformed by applying FUN to them: select using names, indices, a logical vector or a selector function (e.g. is.numeric). Since v1.7 vars is evaluated within the .data environment, permitting expressions on columns e.g. c(col1, col3:coln).
FUN: a single function yielding a result of length NROW(.data) or 1. See also apply.
apply: logical. TRUE (default) will apply FUN to each column selected in vars; FALSE will apply FUN to the subsetted data frame i.e. FUN(get_vars(.data, vars), ...). The latter is useful for collapse functions with data frame or grouped / panel data frame methods, yielding performance gains and enabling grouped transformations. See Examples.
value: a named list of replacements, it will be treated like an evaluated list of column = value arguments.
keep: select columns to preserve using column names, indices or a function (e.g. is.numeric). By default computed columns are added after the preserved ones, unless they are assigned the same name in which case the preserved columns will be replaced in order.
.keep: either one of "all", "used", "unused" or "none" (see mutate), or columns names/indices/function as keep. Note that this does not work well with across() or other expressions supported since v1.9.0. The only sensible option you have there is to supply a character vector of all columns in the final dataset that you want to keep.
.cols: for expressions involving .data, .cols can be used to subset columns, e.g. mtcars |> gby(cyl) |> mtt(broom::augment(lm(mpg ~., .data)), .cols = 1:7). Can pass column names, indices, a logical vector or a selector function (e.g. is.numericr).

Details

The ... arguments to ftransform are tagged vector expressions, which are evaluated in the data frame .data. The tags are matched against names(.data), and for those that match, the values replace the corresponding variable in .data, whereas the others are appended to .data. It is also possible to delete columns by assigning NULL to them, i.e. ftransform(data, colk = NULL) removes colk from the data. Note that names(.data) and the names of the ... arguments are checked for uniqueness beforehand, yielding an error if this is not the case.

Since collapse v1.3.0, is is also possible to pass a single named list to ..., i.e. ftransform(data, newdata). This list will be treated like a list of tagged vector expressions. Note the different behavior: ftransform(data, list(newcol = col1)) is the same as ftransform(data, newcol = col1), whereas ftransform(data, newcol = as.list(col1)) creates a list column. Something like ftransform(data, as.list(col1)) gives an error because the list is not named. See Examples.

The function ftransformv added in v1.3.2 provides a fast replacement for the functions dplyr::mutate_at and dplyr::mutate_if (without the grouping feature) facilitating mutations of groups of columns (dplyr::mutate_all is already accounted for by dapply). See Examples.

The function settransform does all of that by reference, but uses base-R's copy-on modify semantics, which is equivalent to replacing the data with <- (thus it is still memory efficient but the data will have a different memory address afterwards).

The function fcompute(v) works just like ftransform(v), but returns only the changed / computed columns without modifying or appending the data in .data. See Examples.

The function fmutate added in v1.7.0, provides functionality familiar from dplyr 1.0.0 and higher. It evaluates tagged vector expressions sequentially and does operations by groups on a grouped frame (thus it is slower than ftransform if you have many tagged expressions or a grouped data frame). Note however that collapse does not depend on rlang, so things like lambda expressions are not available. Note also that fmutate operates differently on grouped data whether you use .FAST_FUN or base R functions / functions from other packages. With .FAST_FUN (including .OPERATOR_FUN, excluding fhdbetween / fhdwithin / HDW / HDB), fmutate performs an efficient vectorized execution, i.e. the grouping object from the grouped data frame is passed to the g argument of these functions, and for .FAST_STAT_FUN also TRA = "replace_fill" is set (if not overwritten by the user), yielding internal grouped computation by these functions without the need for splitting the data by groups. For base R and other functions, fmutate performs classical split-apply combine computing i.e. the relevant columns of the data are selected and split into groups, the expression is evaluated for each group, and the result is recombined and suitably expanded to match the original data frame. Note that it is not possible to mix vectorized and standard execution in the same expression!! Vectorized execution is performed if any .FAST_FUN or .OPERATOR_FUN is part of the expression, thus a code like mtcars |> gby(cyl) |> fmutate(new = fmin(mpg) / min(mpg)) will be expanded to something like mtcars |> gby(cyl) |> ftransform(new = fmin(mpg, g = GRP(.), TRA = "replace_fill") / min(mpg)) and then executed, i.e. fmin(mpg) will be executed in a vectorized way, and min(mpg) will not be executed by groups at all.

Examples

Run this code


## fmutate() examples ---------------------------------------------------------------

# Please note that expressions are vectorized whenever they contain 'ANY' fast function
mtcars |>
  fgroup_by(cyl, vs, am) |>
  fmutate(mean_mpg = fmean(mpg),                     # Vectorized
          mean_mpg_base = mean(mpg),                 # Non-vectorized
          mpg_cumpr = fcumsum(mpg) / fsum(mpg),      # Vectorized
          mpg_cumpr_base = cumsum(mpg) / sum(mpg),   # Non-vectorized
          mpg_cumpr_mixed = fcumsum(mpg) / sum(mpg)) # Vectorized: division by overall sum

# Using across: here fmean() gets vectorized across both groups and columns (requiring a single
# call to fmean.data.frame which goes to C), whereas weighted.mean needs to be called many times.
mtcars |> fgroup_by(cyl, vs, am) |>
  fmutate(across(disp:qsec, list(mu = fmean, mu2 = weighted.mean), w = wt, .names = "flip"))

# Can do more complex things...
mtcars |> fgroup_by(cyl) |>
  fmutate(res = resid(lm(mpg ~ carb + hp, weights = wt)))

# Since v1.9.0: supports arbitrary expressions returning suitable lists
if (FALSE)  
mtcars |> fgroup_by(cyl) |>
  fmutate(broom::augment(lm(mpg ~ carb + hp, weights = wt)))

# Same thing using across() (supported before 1.9.0)
modelfun <- function(data) broom::augment(lm(mpg ~ carb + hp, data, weights = wt))
mtcars |> fgroup_by(cyl) |>
  fmutate(across(c(mpg, carb, hp, wt), modelfun, .apply = FALSE))



## ftransform() / fcompute() examples: ----------------------------------------------

## ftransform modifies and returns a data.frame
head(ftransform(airquality, Ozone = -Ozone))
head(ftransform(airquality, new = -Ozone, Temp = (Temp-32)/1.8))
head(ftransform(airquality, new = -Ozone, new2 = 1, Temp = NULL))  # Deleting Temp
head(ftransform(airquality, Ozone = NULL, Temp = NULL))            # Deleting columns

# With collapse's grouped and weighted functions, complex operations are done on the fly
head(ftransform(airquality, # Grouped operations by month:
                Ozone_Month_median = fmedian(Ozone, Month, TRA = "fill"),
                Ozone_Month_sd = fsd(Ozone, Month, TRA = "replace"),
                Ozone_Month_centered = fwithin(Ozone, Month)))

# Grouping by month and above/below average temperature in each month
head(ftransform(airquality, Ozone_Month_high_median =
                  fmedian(Ozone, list(Month, Temp > fbetween(Temp, Month)), TRA = "fill")))

## ftransformv can be used to modify multiple columns using a function
head(ftransformv(airquality, 1:3, log))
head(`[<-`(airquality, 1:3, value = lapply(airquality[1:3], log))) # Same thing in base R

head(ftransformv(airquality, 1:3, log, apply = FALSE))
head(`[<-`(airquality, 1:3, value = log(airquality[1:3])))         # Same thing in base R

# Using apply = FALSE yields meaningful performance gains with collapse functions
# This calls fwithin.default, and repeates the grouping by month 3 times:
head(ftransformv(airquality, 1:3, fwithin, Month))

# This calls fwithin.data.frame, and only groups one time -> 5x faster!
head(ftransformv(airquality, 1:3, fwithin, Month, apply = FALSE))

# This also works for grouped and panel data frames (calling fwithin.grouped_df)
airquality |> fgroup_by(Month) |>
  ftransformv(1:3, fwithin, apply = FALSE) |> head()

# But this gives the WRONG result (calling fwithin.default). Need option apply = FALSE!!
airquality |> fgroup_by(Month) |>
  ftransformv(1:3, fwithin) |> head()

# For grouped modification of single columns in a grouped dataset, we can use GRP():
library(magrittr)
airquality |> fgroup_by(Month) %>%
  ftransform(W_Ozone = fwithin(Ozone, GRP(.)),                 # Grouped centering
             sd_Ozone_m = fsd(Ozone, GRP(.), TRA = "replace"), # In-Month standard deviation
             sd_Ozone = fsd(Ozone, TRA = "replace"),           # Overall standard deviation
             sd_Ozone2 = fsd(Ozone, TRA = "fill"),             # Same, overwriting NA's
             sd_Ozone3 = fsd(Ozone)) |> head()                 # Same thing (calling alloc())

## For more complex mutations we can use ftransform with compound pipes
airquality |> fgroup_by(Month) %>%
  ftransform(get_vars(., 1:3) |> fwithin() |> flag(0:2)) |> head()

airquality %>% ftransform(STD(., cols = 1:3) |> replace_na(0)) |> head()

# The list argument feature also allows flexible operations creating multiple new columns
airquality |> # The variance of Wind and Ozone, by month, weighted by temperature:
  ftransform(fvar(list(Wind_var = Wind, Ozone_var = Ozone), Month, Temp, "replace")) |> head()

# Same as above using a grouped data frame (a bit more complex)
airquality |> fgroup_by(Month) %>%
  ftransform(fselect(., Wind, Ozone) |> fvar(Temp, "replace") |> add_stub("_var", FALSE)) |>
  fungroup() |> head()

# This performs 2 different multi-column grouped operations (need c() to make it one list)
ftransform(airquality, c(fmedian(list(Wind_Day_median = Wind,
                                      Ozone_Day_median = Ozone), Day, TRA = "replace"),
                         fsd(list(Wind_Month_sd = Wind,
                                  Ozone_Month_sd = Ozone), Month, TRA = "replace"))) |> head()

## settransform(v) works like ftransform(v) but modifies a data frame in the global environment..
settransform(airquality, Ratio = Ozone / Temp, Ozone = NULL, Temp = NULL)
head(airquality)
rm(airquality)

# Grouped and weighted centering
settransformv(airquality, 1:3, fwithin, Month, Temp, apply = FALSE)
head(airquality)
rm(airquality)

# Suitably lagged first-differences
settransform(airquality, get_vars(airquality, 1:3) |> fdiff() |> flag(0:2))
head(airquality)
rm(airquality)

# Same as above using magrittr::`%<>%`
airquality %<>% ftransform(get_vars(., 1:3) |> fdiff() |> flag(0:2))
head(airquality)
rm(airquality)

# It is also possible to achieve the same thing via a replacement method (if needed)
ftransform(airquality) <- get_vars(airquality, 1:3) |> fdiff() |> flag(0:2)
head(airquality)
rm(airquality)

## fcompute only returns the modified / computed columns
head(fcompute(airquality, Ozone = -Ozone))
head(fcompute(airquality, new = -Ozone, Temp = (Temp-32)/1.8))
head(fcompute(airquality, new = -Ozone, new2 = 1))

# Can preserve existing columns, computed ones are added to the right if names are different
head(fcompute(airquality, new = -Ozone, new2 = 1, keep = 1:3))

# If given same name as preserved columns, preserved columns are replaced in order...
head(fcompute(airquality, Ozone = -Ozone, new = 1, keep = 1:3))

# Same holds for fcomputev
head(fcomputev(iris, is.numeric, log)) # Same as:
iris |> get_vars(is.numeric) |> dapply(log) |> head()

head(fcomputev(iris, is.numeric, log, keep = "Species"))   # Adds in front
head(fcomputev(iris, is.numeric, log, keep = names(iris))) # Preserve order

# Keep a subset of the data, add standardized columns
head(fcomputev(iris, 3:4, STD, apply = FALSE, keep = names(iris)[3:5]))