categories_to_binary: Get binary representation of categorical variables

In anticlust: Subset Partitioning via Anticlustering

Get binary representation of categorical variables

Description

Get binary representation of categorical variables

Usage

categories_to_binary(categories, use_combinations = FALSE)

Arguments

categories	A vector, data.frame or matrix representing one or several categorical variables
use_combinations	Logical, should the output also include columns representing the combination / interaction of the categories (defaults to FALSE).

Details

The conversion of categorical variables to binary variables is done via model.matrix. Since version 0.8.9, each category of a categorical variable is coded by a separate variable ('one hot' encoding). So we do not use 'dummy' coding, which is often used to encode predictors in statistical analysis. Dummy coding uses a reference category that is not explicitly modelled via a separate variable. This implies that there is a different distance to the reference category than among the other categories, which is unwarranted in anticlustering (thanks to Gunnar Klau for noting this). See examples.

This function can be used to include categorical variables as part of the optimization criterion in anticlustering, rather than including them as hard constraints as done when using the argument categories in anticlustering (or fast_anticlustering). This way, categorical variables are treated as numeric variables, which can be useful when there are several categorical variables or when the group sizes are unequal (or both). See examples. Please see the vignette 'Using categorical variables with anticlustering' for more information on this approach.

Since version 0.8.12, this function treats NA in the categorical input variables as a distinct category. This implies that different observations having an NA are treated as having the same value on this variable. If this is not desired, you need to do some other preprocessing of the NAs before calling this function.

Value

A matrix encoding the categorical variable(s) in binary form.

Author(s)

Martin Papenberg [e-mail]

References

Papenberg, M., Wang, C., Diop, M., Bukhari, S. H., Oskotsky, B., Davidson, B. R., Vo, K. C., Liu, B., Irwin, J. C., Combes, A., Gaudilliere, B., Li, J., Stevenson, D. K., Klau, G. W., Giudice, L. C., Sirota, M., & Oskotsky, T. T. (2025). Anticlustering for sample allocation to minimize batch effects. Cell Reports Methods, 5(8), 101137. https://doi.org/10.1016/j.crmeth.2025.101137

Examples

# Illustrate why dummy encoding is not appropriate for anticlustering.
# Use 3 elements of the iris data set, with different levels of 'Species'
input <- iris[c(1, 51, 101), "Species", drop = FALSE]
input
# Default dummy encoding: 
(dummy <- model.matrix(~ . , input))
dist(dummy) 
# Distance between versicolor and virginica is larger than among setosa and 
# versicolor, and setosa and virginica. This would bias the anticlustering 
# computation.
(one_hot <- categories_to_binary(input))
dist(one_hot) # all the same distances 
dist(one_hot)^2 # or use squared Euclidean distance (corresponds to Manhattan distance in this case)
dist(one_hot, method = "manhattan")

# We could also get one-hot encoding directly via model.matrix(), which is 
# used in categories_to_binary():
model.matrix(~ . -1, input) # suppresses the 'intercept'

# Use Schaper data set for anticlustering example
data(schaper2019)
features <- schaper2019[, 3:6]
K <- 3
N <- nrow(features) 

# - Generate data input for k-means anticlustering -
# We conduct k-plus anticlustering by first generating k-plus variables, 
# and also include the categorical variable as "numeric" input for the 
# k-means optimization (rather than as input for the argument categories)

input_data <- cbind(
  kplus_moment_variables(features, T = 2), 
  categories_to_binary(schaper2019$room) 
)

kplus_groups <- anticlustering(
  input_data, 
  K = K,
  objective = "variance",
  method = "local-maximum", 
  repetitions = 10
)
mean_sd_tab(features, kplus_groups)
table(kplus_groups, schaper2019$room) # argument categories was not used!

 

anticlust documentation built on April 15, 2026, 5:06 p.m.