mdl

mdl implements an opinionated and performant reimagining of model matrices. The package supplies one function, mdl::mtrx() (read: “model matrix”), that takes in a formula and data frame and outputs a numeric matrix. Compared to its base R friend model.matrix(), it’s really fast.

This package is highly experimental. Interpret results with caution!

Installation

You can install the development version of mdl like so:

# install.packages("mdl")
pak::pak("simonpcouch/mdl")

Example

The output of mdl::mtrx() looks a lot like that from model.matrix():

# convert to factor to demonstrate dummy variable creations
mtcars$cyl <- as.factor(mtcars$cyl)

head(
  mdl::mtrx(mpg ~ ., mtcars)
)
#>   (Intercept) cyl6 cyl8 disp  hp drat    wt  qsec vs am gear carb
#> 1           1    1    0  160 110 3.90 2.620 16.46  0  1    4    4
#> 2           1    1    0  160 110 3.90 2.875 17.02  0  1    4    4
#> 3           1    0    0  108  93 3.85 2.320 18.61  1  1    4    1
#> 4           1    1    0  258 110 3.08 3.215 19.44  1  0    3    1
#> 5           1    0    1  360 175 3.15 3.440 17.02  0  0    3    2
#> 6           1    1    0  225 105 2.76 3.460 20.22  1  0    3    1

Compared to model.matrix(), mdl::mtrx() is sort of a glorified as.matrix() data frame method. More specifically:

• Does not accept formulae with inlined functions (like - or *).
• Never drops rows (and thus doesn’t accept an na.action).
• Assumes that factors levels are encoded as they’re intended (i.e. drop.unused.levels and xlev are not accepted).

It’s quite a bit faster for smaller data sets:

bench::mark(
  mdl::mtrx(mpg ~ ., mtcars),
  model.matrix(mpg ~ ., mtcars),
  check = FALSE
)
#> # A tibble: 2 × 6
#>   expression                         min   median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                    <bch:tm> <bch:tm>     <dbl> <bch:byt>    <dbl>
#> 1 mdl::mtrx(mpg ~ ., mtcars)      23.1µs     26µs    37187.    3.32KB     18.6
#> 2 model.matrix(mpg ~ ., mtcars)  270.2µs    293µs     3337.  494.24KB     31.9

The factor of speedup isn’t so drastic for larger datasets and datasets with more factors, but it is still quite substantial:

for (p in c("vs", "am", "gear", "carb")) {
  mtcars[[p]] <- as.factor(mtcars[[p]])
}

bench::mark(
  mdl::mtrx(mpg ~ ., mtcars[rep(1:32, 1e5), ]),
  model.matrix(mpg ~ ., mtcars[rep(1:32, 1e5), ]),
  check = FALSE
)
#> Warning: Some expressions had a GC in every iteration; so filtering is
#> disabled.
#> # A tibble: 2 × 6
#>   expression                             min median `itr/sec` mem_alloc `gc/sec`
#>   <bch:expr>                           <bch> <bch:>     <dbl> <bch:byt>    <dbl>
#> 1 mdl::mtrx(mpg ~ ., mtcars[rep(1:32,… 1.43s  1.43s     0.701  803.01MB    0.701
#> 2 model.matrix(mpg ~ ., mtcars[rep(1:… 2.01s  2.01s     0.497    1.86GB    1.99

Check out this article for more detailed benchmarks.