Modelling Functions that Work with the Pipe

Functions for modelling that help you seamlessly integrate modelling into a pipeline of data manipulation and visualisation.

The modelr package provides functions that help you create elegant pipelines when modelling. Its design follows Hadley Wickham's tidy tool manifesto.

Installation and Documentation

You can install modelr from github with:

devtools::install_github("tidyverse/modelr")

Alternatively, modelr is available as part of the tidyverse package which can be installed via:

install.packages("tidyverse")

Note that unlike the core tidyverse packages, modelr would not be loaded via library(tidyverse). Instead, you can load it explicitly:

library(modelr)

Full documentation is available in R for Data Science, mostly in the Model basics chapter.

Main Features

Partitioning and sampling

The resample class stores a "pointer" to the original dataset and a vector of row indices. resample can be turned into a dataframe by calling as.data.frame. The indices can be extracted using as.integer:

# a subsample of the first ten rows in the data frame
rs <- resample(mtcars, 1:10)
as.data.frame(rs)
#>                    mpg cyl  disp  hp drat    wt  qsec vs am gear carb
#> Mazda RX4         21.0   6 160.0 110 3.90 2.620 16.46  0  1    4    4
#> Mazda RX4 Wag     21.0   6 160.0 110 3.90 2.875 17.02  0  1    4    4
#> Datsun 710        22.8   4 108.0  93 3.85 2.320 18.61  1  1    4    1
#> Hornet 4 Drive    21.4   6 258.0 110 3.08 3.215 19.44  1  0    3    1
#> Hornet Sportabout 18.7   8 360.0 175 3.15 3.440 17.02  0  0    3    2
#> Valiant           18.1   6 225.0 105 2.76 3.460 20.22  1  0    3    1
#> Duster 360        14.3   8 360.0 245 3.21 3.570 15.84  0  0    3    4
#> Merc 240D         24.4   4 146.7  62 3.69 3.190 20.00  1  0    4    2
#> Merc 230          22.8   4 140.8  95 3.92 3.150 22.90  1  0    4    2
#> Merc 280          19.2   6 167.6 123 3.92 3.440 18.30  1  0    4    4
as.integer(rs)
#>  [1]  1  2  3  4  5  6  7  8  9 10

The class can be utilized in generating an exclusive partitioning of a data frame:

# generate a 30% testing partition and a 70% training partition
ex <- resample_partition(mtcars, c(test = 0.3, train = 0.7))
lapply(ex, dim)
#> $test
#> [1]  9 11
#> 
#> $train
#> [1] 23 11

modelr offers several resampling methods that result in a list of resample objects (organized in a data frame):

# bootstrap
boot <- bootstrap(mtcars, 100)
# k-fold cross-validation
cv1 <- crossv_kfold(mtcars, 5)
# Monte Carlo cross-validation
cv2 <- crossv_mc(mtcars, 100)
 
dim(boot$strap[[1]])
#> [1] 32 11
dim(cv1$train[[1]])
#> [1] 25 11
dim(cv1$test[[1]])
#> [1]  7 11
dim(cv2$train[[1]])
#> [1] 25 11
dim(cv2$test[[1]])
#> [1]  7 11

Model quality metrics

modelr includes several often-used model quality metrics:

mod <- lm(mpg ~ wt, data = mtcars)
rmse(mod, mtcars)
#> [1] 2.949163
rsquare(mod, mtcars)
#> [1] 0.7528328
mae(mod, mtcars)
#> [1] 2.340642
qae(mod, mtcars)
#>        5%       25%       50%       75%       95% 
#> 0.1784985 1.0005640 2.0946199 3.2696108 6.1794815

Interacting with models

A set of functions let you seamlessly add predictions and residuals as additional columns to an existing data frame:

df <- tibble::data_frame(
  x = sort(runif(100)),
  y = 5 * x + 0.5 * x ^ 2 + 3 + rnorm(length(x))
)
 
mod <- lm(y ~ x, data = df)
df %>% add_predictions(mod)
#> # A tibble: 100 x 3
#>              x        y     pred
#>          <dbl>    <dbl>    <dbl>
#>  1 0.007764089 3.434399 3.065858
#>  2 0.010139684 3.924469 3.078856
#>  3 0.020426476 3.574665 3.135142
#>  4 0.054073586 1.786603 3.319247
#>  5 0.075632210 2.544859 3.437208
#>  6 0.099335418 2.855342 3.566903
#>  7 0.122253743 3.571305 3.692304
#>  8 0.138008688 3.867327 3.778509
#>  9 0.152028988 3.671963 3.855223
#> 10 0.152634759 3.860819 3.858538
#> # ... with 90 more rows
df %>% add_residuals(mod)
#> # A tibble: 100 x 3
#>              x        y        resid
#>          <dbl>    <dbl>        <dbl>
#>  1 0.007764089 3.434399  0.368540719
#>  2 0.010139684 3.924469  0.845612896
#>  3 0.020426476 3.574665  0.439523271
#>  4 0.054073586 1.786603 -1.532643175
#>  5 0.075632210 2.544859 -0.892348715
#>  6 0.099335418 2.855342 -0.711561273
#>  7 0.122253743 3.571305 -0.120998985
#>  8 0.138008688 3.867327  0.088817647
#>  9 0.152028988 3.671963 -0.183260285
#> 10 0.152634759 3.860819  0.002281414
#> # ... with 90 more rows

For visualization purposes it is often useful to use an evenly spaced grid of points from the data:

data_grid(mtcars, wt = seq_range(wt, 10), cyl, vs)
#> # A tibble: 60 x 3
#>          wt   cyl    vs
#>       <dbl> <dbl> <dbl>
#>  1 1.513000     4     0
#>  2 1.513000     4     1
#>  3 1.513000     6     0
#>  4 1.513000     6     1
#>  5 1.513000     8     0
#>  6 1.513000     8     1
#>  7 1.947556     4     0
#>  8 1.947556     4     1
#>  9 1.947556     6     0
#> 10 1.947556     6     1
#> # ... with 50 more rows
 
# For continuous variables, seq_range is useful
mtcars_mod <- lm(mpg ~ wt + cyl + vs, data = mtcars)
data_grid(mtcars, wt = seq_range(wt, 10), cyl, vs) %>% add_predictions(mtcars_mod)
#> # A tibble: 60 x 4
#>          wt   cyl    vs     pred
#>       <dbl> <dbl> <dbl>    <dbl>
#>  1 1.513000     4     0 28.37790
#>  2 1.513000     4     1 28.90207
#>  3 1.513000     6     0 25.64969
#>  4 1.513000     6     1 26.17386
#>  5 1.513000     8     0 22.92148
#>  6 1.513000     8     1 23.44566
#>  7 1.947556     4     0 26.96717
#>  8 1.947556     4     1 27.49134
#>  9 1.947556     6     0 24.23896
#> 10 1.947556     6     1 24.76314
#> # ... with 50 more rows