encodeR

This package seeks to provide a convenient set of functions that allow for the encoding of categorical features in potentially more informative ways when compared to other, more standard methods. The user will feed as input a training and testing dataset with categorical features, and the resulting data frames returned will be preprocessed with a specific encoding of the categorical features. At a high level, this package automates the preprocessing of categorical features in ways that exploit particular correlations between the different categories and the data without increasing the dimension of the dataset, like in one hot encoding. Thus, through the more deliberate handling of these categorical features, higher model performance can possibly be achieved.

Please read this vignette for a hands on guide on how to use encodeR.

Features

This package contains four functions, each that accept two tibbles representing the train and test sets. Depending on the method, the functions will also require additional arguments depending on how the encodings are calculated for each category. For now, the package aims to support binary classification and regression problems.

onehot_encoder(): the standard one-hot encoding of categorical features, which will create K-1 columns of 0/1 indicator variables.
frequency_encoder(): calculates encodings based off the observed frequency of each category in the training set.
target_encoder(): calculates encodings by computing the average observed response per each category.
conjugate_encoder(): calculates encodings based off Bayes rule using conjugate priors and the mean of the posterior distribution. The original paper for this method can be found here.

Where encodeR Fits in The R Ecosystem

There are some packages in R that include different, more sophisticated kinds of encoding methods. The well known framework H20 has a function for target encoding, and the recipes package has the ability to one hot encode. The package cattonum also contains many kinds of encoding schemes such as frequency encoding, target encoding, and one hot encoding. Based on research, there does not exist a package in R that implements conjugate encoding.

The problem with the R ecosystem for categorical encoding is that there is not one package that contains all of the most popular encoders. This results in users having to import many different packages to experiment with different encoders, each with their own syntax and interface. This package hopes to give users one, coherent framework for encoding categorical features in R. Furthermore, methods that have not been packaged in R like conjugate encoding will directly add something new to the R ecosystem.

Installation

Currently, this package is not yet on CRAN. However, you can install the latest development version from GitHub with:

# install.packages("devtools")
devtools::install_github("UBC-MDS/encodeR")

Dependencies

R (>= 3.4)
dplyr
readr
rlang
tidyr
purrr
fastDummies
magrittr

Examples

This package can allow one to fit many different kinds of encodings for categorical features easily. For example, one can fit a conjugate encoding of some categorical variables by:

library(encodeR)
library(tidyverse)
#> ── Attaching packages ────────────────────────────────────────────────────────────────────────────────────────────── tidyverse 1.3.0 ──
#> ✓ ggplot2 3.3.0     ✓ purrr   0.3.3
#> ✓ tibble  2.1.3     ✓ dplyr   0.8.5
#> ✓ tidyr   1.0.2     ✓ stringr 1.4.0
#> ✓ readr   1.3.1     ✓ forcats 0.4.0
#> ── Conflicts ───────────────────────────────────────────────────────────────────────────────────────────────── tidyverse_conflicts() ──
#> x dplyr::filter() masks stats::filter()
#> x dplyr::lag()    masks stats::lag()

my_data <- tibble(
  fruit = c("Apple", "Orange", "Apple", "Apple", "Banana"),
  color = c("Red", "Blue", "Orange", "Red", "Red"),
  target = c(1, 0, 1, 1, 1)
)

conjugate_encoder(
  my_data,
  y = my_data$target,
  cat_columns = c("fruit", "color"),
  prior_params = list(alpha = 3, beta = 10), 
  objective = "binary"
  )
#> Joining, by = "fruit"
#> Joining, by = "color"
#> [[1]]
#> # A tibble: 5 x 3
#>   target fruit_encoded color_encoded
#>    <dbl>         <dbl>         <dbl>
#> 1      1         0.333         0.333
#> 2      0         0.167         0.167
#> 3      1         0.333         0.222
#> 4      1         0.333         0.333
#> 5      1         0.222         0.333

This package can also fit regression data sets, and automatically join the learned encodings on a held out test set if the user wants to:

my_data_reg <- tibble(
  fruit = c("Apple", "Orange", "Apple", "Apple", "Banana", "Orange", "Apple"),
  color = c("Red", "Blue", "Orange", "Red", "Red", "Blue", "Green"),
  target = c(3.5, 10, 10.912, 3.14159, 10, 15, 1000)
)

train <- my_data_reg[1:5, ]
test <- my_data_reg[6:7, ]

encodings <- target_encoder(
  train,
  test,
  y = train$target,
  cat_columns = c("fruit", "color"),
  prior = 0.8, 
  objective = "regression" 
)

encodings$train
#> # A tibble: 5 x 3
#>   fruit color target
#>   <dbl> <dbl>  <dbl>
#> 1  6.20  5.96   3.5 
#> 2  8.89  8.89  10   
#> 3  6.20  9.40  10.9 
#> 4  6.20  5.96   3.14
#> 5  8.89  5.96  10
encodings$test
#> # A tibble: 2 x 3
#>   fruit color target
#>   <dbl> <dbl>  <dbl>
#> 1  8.89  8.89     15
#> 2  6.20  7.51   1000