Lumps the levels of an unordered categorical variable so that every resulting
level meets the sample-size threshold, choosing the combination that
preserves the most mutual information with the original variable. An optional
preference graph restricts which levels may be combined.
Usage
lump_nominal(
data,
threshold,
adj_matrix = NULL,
verbose = FALSE,
alternative_metric = c("mutual information", "bin count", "surplus"),
level_namer = default_level_namer
)Arguments
- data
Factor or character vector of the categorical data.
- threshold
The minimum number of samples each lumped level should contain.
- adj_matrix
Adjancency matrix of the preference graph. Default: a complete graph, allowing all lumpings.
- verbose
Logical value dictating if values should be printed. Default:
FALSE.- alternative_metric
The metric that should be optimised for, if it is different from the default, the mutual information. For an explanation of the metrics see
vignette("metrics").- level_namer
Function that takes a character vector of the original levels in a lump and returns the name of the new lumped level. Default: concatenating the original levels with a "+" in between.
See also
maximum_mutual_information_nominal() for the underlying algorithm that this function wraps.
lump_hierarchical() for a version of this function that can take advantage of hierarchical structure in the data to speed up the execution time.
lump_nominal_heuristic() to approximate this function when the runtime becomes infeasible.
lump_nominal_supervised() for a supervised version of this function.
Examples
m <- 5
n <- 50
q <- 10
data <- sample(LETTERS[1:m], n, replace = TRUE)
lump_nominal(data, q)
#> [1] E A+D E A+D A+D E E E B C C E E A+D E B A+D B C
#> [20] A+D C E A+D B E B A+D A+D A+D B C A+D A+D B E A+D A+D C
#> [39] A+D A+D C B C B B C A+D B A+D C
#> Levels: A+D B C E