Bringing some speed to `concat.split`

concat.split.DT.R

#' Split concatenated cells in a \code{data.frame} or a \code{data.table}
#' 
#' A variation of the \code{concat.split} family of functions designed for
#' large rectangular datasets.
#' 
#' While the general \code{concat.split} functions are able to handle
#' "unbalanced" datasets (for example, where the number of fields in a given
#' column might differ from row to row) because of the nature of \code{fread}
#' from the "data.table" package, this function does not support such data
#' types.
#' 
#' @param dataset The input \code{data.frame} or \code{data.table}.
#' @param splitcols The columns that need to be split up.
#' @param sep The character that serves as a delimiter within the columns that
#' need to be split up.
#' @param drop Logical. Should the original columns be dropped? Defaults to
#' \code{TRUE}.
#' @param dotsub The character that should be substituted as a delimiter
#' \emph{if \code{sep = "."}}. \code{fread} does not seem to work nicely with
#' \code{sep = "."}, so it needs to be substituted. By default, this function
#' will substitute \code{"."} with \code{"|"}.
#' @return A \code{data.table}.
#' @author Ananda Mahto
#' @references \url{http://stackoverflow.com/a/19231054/1270695}
#' @examples
#' 
#' small_file <- system.file("concatDT.csv", package = "SOfun")
#' small_data <- read.csv(small_file)
#' dim(small_data)
#' head(small_data)
#' out <- concat.split.DT(small_data,
#'                        splitcols = c("VARIABLE", "VAR2", "VAR3", "VAR4"),
#'                        sep = "_", drop = TRUE)
#' out
#' 
#' \dontrun{
#' ## Make a much bigger dataset
#' big_data <- small_data[rep(rownames(small_data),
#'                            1500000/nrow(small_data)), ]
#' dim(big_data)
#' system.time(big_out <- concat.split.DT(big_data,
#'                                        splitcols = c("VARIABLE", "VAR2",
#'                                                      "VAR3", "VAR4"),
#'                                        sep = "_", drop = TRUE))
#' big_out
#' }
#' 
#' @export concat.split.DT
concat.split.DT <- function(dataset, splitcols, sep, drop = TRUE, dotsub = "|") {
  require(data.table)
  if (is.numeric(splitcols)) splitcols <- names(dataset)[splitcols]
  if (!is.data.table(dataset)) dataset <- data.table(dataset)
  if (sep == ".") {
    dataset[, (splitcols) := gsub(".", dotsub, get(splitcols), fixed = TRUE)]
    sep <- dotsub
  }

  Splits <- do.call(cbind, lapply(splitcols, function(Z) {
    x <- tempfile()
    if (!is.character(dataset[[Z]])) writeLines(as.character(dataset[[Z]]), x)
    else writeLines(dataset[[Z]], x)
    Split <- fread(x, sep = sep, header = FALSE)
    setnames(Split, paste(Z, seq_along(Split), sep = "_"))
    Split
  }))
  
  final <- cbind(dataset, Splits)
  if (isTRUE(drop)) final <- final[, setdiff(names(final), splitcols), with = FALSE]
  final
}

Get some medium-sized sample data here. Make it large with:

out <- mydata[rep(rownames(mydata), 1500000/nrow(mydata)), ]

Try the function:

concat.split.DT(out, "VARIABLE", "_")

@mrdwab, looks good. Here are some suggestions:

1. as.data.table(.) is faster than data.table(.).
2. Your sep == "." if-statement will modify dataset by reference, if the input dataset is already a data.table. Is this intended?
3. Use of tempfile() is fine. A text=. argument would be much nicer, I agree. I'll write to Matt about this.
4. A bit more speed-up can be obtained by using set instead of cbind.

Here's a proof of concept. This function ran in 0.8 seconds as opposed to 2.1 secs on the data out. But it'd be better to benchmark it on bigger data sets (especially with length(splitcols) > 1).

concat.split.DT.mod <- function(dataset, splitcols, sep, drop = TRUE, dotsub = "|") {
    if (!require(data.table)) stop("data.table package not found")
    if (!length(splitcols)) {
        warning("No columns to split by, returning dataset as such.")
        return(dataset)
    }
    if (is.numeric(splitcols)) splitcols <- names(dataset)[splitcols]
    if (!is.data.table(dataset)) dataset <- as.data.table(dataset) else dataset <- copy(dataset)
    if (sep == ".") {
        dataset[, (splitcols) := gsub(".", dotsub, get(splitcols), fixed = TRUE)]
        sep <- dotsub
    }
    write_fread <- function(Z) {
        x = tempfile()
        if (!is.character(dataset[[Z]])) writeLines(as.character(dataset[[Z]]), x)
        else writeLines(dataset[[Z]], x)
        fread(x, sep = sep, header = FALSE)
    }
    split_names <- function(Z, Split) {
        paste(Z, seq_along(Split), sep = "_")
    }

    # treat splitcols[1L] first
    Splits = write_fread(splitcols[1L])
    setnames(Splits, split_names(splitcols[1L], Splits))
    for (i in seq_along(splitcols)[-1L]) {
        tmp = write_fread(splitcols[i])
        set(Splits, i = NULL, j = split_names(splitcols[i], tmp), value = tmp)
    }
    set(dataset, i = NULL, j = names(Splits), value = Splits)
    if (isTRUE(drop)) set(dataset, i = NULL, j = splitcols, value = NULL)
    dataset
}

However, one check you'll have to do, if you use set, is to make sure that the new columns you'll create with set don't already exist in the data.table. If so, they'll be overwritten. cbind on the other hand binds duplicate names as such. But it'll make a copy - for every binding.

So, if you can take care of this issue, I'd suggest working on the copy (either with as.data.table(.) or copy(.) as shown above) and using set.

HTH

This part:

    Splits = write_fread(splitcols[1L])
    setnames(Splits, split_names(splitcols[1L], Splits))
    for (i in seq_along(splitcols)[-1L]) {
        tmp = write_fread(splitcols[i])
        set(Splits, i = NULL, j = split_names(splitcols[i], tmp), value = tmp)
    }
    set(dataset, i = NULL, j = names(Splits), value = Splits)
    if (isTRUE(drop)) set(dataset, i = NULL, j = splitcols, value = NULL)
    dataset

can be written directly with dataset instead as:

    for (i in seq_along(splitcols)) {
        Split = write_fread(splitcols[i])
        set(dataset, i = NULL, j = split_names(splitcols[i], Split), value = Split)
    }
    if (isTRUE(drop)) set(dataset, i = NULL, j = splitcols, value = NULL)
    dataset

That'll avoid the second set and add by reference directly yo dataset.
HTH