devtools::install_github("https://github.com/Zelazny7/onehot")
set.seed(100)
test <- data.frame(
factor = factor(sample(c(NA, letters[1:3]), 100, T)),
integer = as.integer(runif(100) * 10),
real = rnorm(100),
logical = sample(c(T, F), 100, T),
character = sample(letters, 100, T),
stringsAsFactors = FALSE)
head(test)
## factor integer real logical character
## 1 a 3 -0.3329234 FALSE f
## 2 a 3 1.3631137 TRUE t
## 3 b 0 -0.4691473 TRUE h
## 4 <NA> 3 0.8428756 TRUE k
## 5 a 5 -1.4579937 FALSE k
## 6 a 6 -0.4003059 FALSE l
A onehot object contains information about the data.frame. This is used to transform a data.frame into a onehot encoded matrix. It should be saved to transform future datasets into the same exact layout.
library(onehot)
## Loading required package: Matrix
encoder <- onehot(test)
## Warning: Variables excluded for having levels > max_levels: character
## printe a summary
encoder
## Onehot Specification
## |- 1 Factors => 4 Indicators
## |- 3 Numerics => (NA <- -999)
The onehot object has a predict method which may be used to transform a
data.frame. Factors are onehot encoded. Character variables are skipped.
However calling predict with stringsAsFactors=TRUE will convert
character vectors to factors first.
train_data <- predict(encoder, test)
head(train_data)
## factor_a factor_b factor_c factor_NA integer real logical
## [1,] 1 0 0 0 3 -0.3329234 0
## [2,] 1 0 0 0 3 1.3631137 1
## [3,] 0 1 0 0 0 -0.4691473 1
## [4,] 0 0 0 1 3 0.8428756 1
## [5,] 1 0 0 0 5 -1.4579937 0
## [6,] 1 0 0 0 6 -0.4003059 0
add_NA_factors=TRUE (the default) will create an indicator column for
every factor column. Having NAs as a factor level will result in an
indicator column being created without using this option.
encoder <- onehot(test, add_NA_factors=TRUE)
## Warning: Variables excluded for having levels > max_levels: character
train_data <- predict(encoder, test)
head(train_data)
## factor_a factor_b factor_c factor_NA integer real logical
## [1,] 1 0 0 0 3 -0.3329234 0
## [2,] 1 0 0 0 3 1.3631137 1
## [3,] 0 1 0 0 0 -0.4691473 1
## [4,] 0 0 0 1 3 0.8428756 1
## [5,] 1 0 0 0 5 -1.4579937 0
## [6,] 1 0 0 0 6 -0.4003059 0
The sentinel=VALUE argument will replace all numeric NAs with the
provided value. Some ML algorithms such as randomForest and xgboost
do not handle NA values. However, by using sentinel values such
algorithms are usually able to separate them with enough decision-tree
splits. The default value is -999
onehot also provides support for predicting sparse, column compressed
matrices from the Matrix package:
encoder <- onehot(test)
## Warning: Variables excluded for having levels > max_levels: character
train_data <- predict(encoder, test, sparse=TRUE)
head(train_data)
## 6 x 7 sparse Matrix of class "dgCMatrix"
## factor_a factor_b factor_c factor_NA integer real logical
## [1,] 1 . . . 3 -0.3329234 .
## [2,] 1 . . . 3 1.3631137 1
## [3,] . 1 . . . -0.4691473 1
## [4,] . . . 1 3 0.8428756 1
## [5,] 1 . . . 5 -1.4579937 .
## [6,] 1 . . . 6 -0.4003059 .