Manage data columns with dict.table"
In container: Extending Base 'R' Lists

require(container)
require(dplyr)
library(microbenchmark)
library(ggplot2)
library(data.table)
library(tibble)


knitr::opts_chunk$set(
  comment = "#",
  prompt = F,
  tidy = FALSE,
  cache = FALSE,
  collapse = T,
  fig.width = 7
)

old <- options(width = 100L)

Motivation

The dplyr functions select and mutate are widely used to manage data data.frame (or tibble) columns. They cover a wide range of use cases and are applied in quick data exploration as well as in data analysis pipelines. On the other hand, when implementing critical code or building R packages, developers may revert to base R to minimize errors and code dependencies. At least, both mutate and select may require additional checking, for example, to catch column name clashes.

The container package in parts was developed to close this gap. With version 1.0.0, it provides dict.table, which can be considered a data.table with an extended set of functions to add, extract, remove and replace data columns with minimal required additional checking, hopefully resulting in lean and robust code.

This vignette compares basic dplyr and dict.table data column operations and at the end shows that both frameworks can be easily combined.

Column operations

To keep matters simple, we use a tiny data set.

library(container)
library(dplyr)

data <- dict.table(x = c(0.2, 0.5), y = letters[1:2])
data

Add

Let's add columns using mutate.

data %>%
    mutate(ID = 1:2, z = 1)

For someone not familar with the tidyverse, this code block might read somewhat odd as the column is added and not mutated. To add a column via dict.table use add.

data %>%
    add(ID = 1:2, z = 1)

The intend to add a column thus is stated more clearly. Next, instead of ID, let's add another numeric column y, which happens to "name-clash" with the already existing column.

data %>%
    mutate(y = 1)

Of course, the initial y-column has been overwritten. While this was easy to see here, it may not if the data has a lot of columns or if column names are created dynamically during runtime. To catch this, usually some overhead is required.

if ("y" %in% colnames(data)) {
    stop("column y already exists")
} else {
    data %>%
        mutate(y = 1)
}

Let's see the dict.table-operation in comparison.

data %>%
    add(y = 1)

The name clash is caught by default and therefore requires no additional checking.

Modify

If the intend was indeed to overwrite the value, the dict.table function replace_at can be used.

data %>%
    replace_at(y = 1)

# or programmatically
data %>%
    replace_at("y", 1)

As we saw above, if a column does not exist, mutate silently creates it for you. If this is not what you want, which means, you want to make sure something is overwritten, again, a workaround is needed.

if ("ID" %in% colnames(data)) {
    data %>%
        mutate(ID = 1:2)
} else {
    stop("column ID not in data.frame")
}

Once again, the workaround is already "built-in" in the dict.table-framework,

data %>%
    replace_at(ID = 1:2)

that is, replace_at expects the column to exist.

If we were to paraphrase the intend of the mutate function, it probably would be something like "Replace a column or, if it does not exist, add it.". As you may already have guessed, this can also be expressed within the dict.table-framework.

data %>%
    replace_at(ID = 1:2, .add = TRUE)

Remove

A common tidyverse approach to remove a column is based on the select function. One corresponding dict.table-function is delete.

data %>%
  select(-"y")

data %>%
    delete_at("y")

Let's see what happens if the column does not exist in the first place.

data %>%
    select(-"ID")

data %>%
    delete_at("ID")

So in this case, both frameworks will complain. Now assume we want the column to be removed if it exist but otherwise silently ignore the command, for example:

if ("ID" %in% colnames(data)) {
    data %>%
        select(-"ID")
}

The dict.table provides a straight-forward solution via the discard function:

data %>%
    discard_at("ID")

Benchmark

To compare the performance of both frameworks, we benchmark some column operations using the standard 'cars' data set. As a hallmark reference we use data.table.

library(microbenchmark)
library(ggplot2)
library(data.table)
library(tibble)

data = cars
head(cars)

For the benchmark, we add, replace and finally delete a column.

bm <- microbenchmark(control = list(order="inorder"), times = 100,

    dict.table =
        as.dict.table(data) %>%
        add(time = .[["dist"]] / .[["speed"]]) %>%
        replace_at(dist = 0) %>%
        delete_at("speed"),

    `data.table[` =
        as.data.table(data)[
        ][, time := dist / speed
        ][, dist := 0
        ][, speed := NULL],

    dplyr =
        as_tibble(data) %>%
        mutate(time = dist / speed) %>%
        mutate(dist = 0) %>%
        select(-speed)
)
autoplot(bm) + theme_bw()

While dict.table and data.table performed nearly the same there is some distance to dplyr (about 10x). Let's examine each operation in more detail.

data = cars
bm <- microbenchmark(control = list(order="inorder"), times = 100,

    dit <- as.dict.table(data),
    dit <- add(dit, time = dit[["dist"]] / dit[["speed"]]),
    dit <- replace_at(dit, dist = 0),
    dit <- delete_at(dit, "speed"),

    dat <-  as.data.table(data),
    dat[, time := dist / speed],
    dat[, dist := 0],
    dat[, speed := NULL],

    tbl <- as_tibble(data),
    tbl <- mutate(tbl, time = dist / speed),
    tbl <- mutate(tbl, dist = 0),
    tbl <- select(tbl, -speed)
)
autoplot(bm) + theme_bw()

Apparently, the mutate and select operations are the slowest in comparison, which for the most part should be a result of these functions providing non-standard evaluation (NSE) and generally a wide range of ways to specify the desired operation. Unsurprisingly such flexibility comes at a cost.

Since the data.table expressions also involve NSE terms and some overhead, in this benchmark the dict.table performs even best. Having said that, of course, the data.table code can be further improved by avoiding the overhead and instead use reference semantics via the data.table built-in set function.

data = cars
bm <- microbenchmark(control = list(order="inorder"), times = 100,

    dict.table =
        as.dict.table(data) %>%
        add(time = dit[["dist"]] / dit[["speed"]]) %>%
        replace_at(dist = 0) %>%
        delete_at("speed"),

    ref_dict.table =
        as.dict.table(data) %>%
        ref_add(time = .[["dist"]] / .[["speed"]]) %>%
        ref_replace_at(dist = 0) %>%
        ref_delete_at("speed"),


    `data.table[` =
        as.data.table(data)[
        ][, time := dist / speed
        ][, dist := 0
        ][, speed := NULL],

    set_data.table =
        as.data.table(data) %>%
        set(j = "ID", value = .[["dist"]] / .[["speed"]]) %>%
        set(j = "dist", value = 0) %>%
        set(j = "speed", value = NULL)
)

autoplot(bm) + theme_bw()

This puts things back into perspective. We also provided a dict.table version using reference semantic, which is also built-in and results in a slight speed improvement over the standard version. As a result, data.table remains the way to go when speed is key.

Combine dplyr and dict.table

Since a dict.table is fully compatible with dplyr and data.table, all of the presented frameworks can be easily combined in any order.

res = data %>%
    as.dict.table %>%
    .[, time := dist / speed] %>%   # data.table
    replace_at(dist = 0) %>%        # container
    select(-speed)                  # dplyr

Summary

In critical code, it is usually of high priority to avoid unintended data column operations. For this, usually additional code is required to check for the existence or absence of columns. The dict.table framework provides a set of column operations with built-in checking, thereby yielding safer and leaner code out of the box and ultimately freeing the developer from writing some annoying checks over and over again.