Create k-Fold Partitioning — create

Adaptation of the rsample::vfold_cv() function and its utilities. Modified for simplicity and to accommodate two-level stratification. Up to 2 levels of stratification can be specified through the breaks parameter:

No stratification: breaks = NULL
One level stratification: breaks is a list(1), where the name of the list element specifies the stratification variable and the value of the element specifies the stratification.
Two level stratification: breaks is a list(2), where the list names specify the stratification variables and the corresponding values specify the stratification.

Usage

create_kfold(data, k = 10L, repeats = 1L, breaks = NULL, ...)

is.k_split(x)

analysis(object, i = NULL)

assessment(object, i = NULL)

Arguments

data: A data.frame containing the data to be subset.
k: integer(1). The number of partitions (folds) to create.
repeats: integer(1)The number of times to repeat the k-fold partitioning.
breaks: Stratification control. Either NULL or a named list. If NULL, no stratification is performed. See Details.
...: For eventual extensibility. Parameters to be passed to internal stratification machinery but currently limited to a depth argument. The number of stratification bins is based on min(5, floor(n / depth)), where n = length(x).
x: An R object to test.
object: A k_split object.
i: integer(1) or NULL. If an integer, the split corresponding to the analysis or assessment data to be retrieved, otherwise a list of all data splits.

Value

A k_split object. Element data contains the original data. Element splits is a tibble where each row of corresponds to an individual split. The split column contains lists named either "analysis" and "assessment" containing the indices of data to be used for each fold and category. Columns fold and .repeat provide fold and repeat indices for each corresponding split.

is.k_split(): Logical. TRUE if x inherits from class k_split.

analysis(): A list of data frames corresponding to the analysis indices.

assessment(): A list of data frames corresponding to the assessment indices.

Details

For stratification variables that are factor, character, or discrete with 5 or fewer unique levels, the stratification structure should be set to NULL. For example, if status is a binary variable, breaks = list(status = NULL).

If the stratification variable has more than 5 unique levels, the stratification structure can be specified as either the number of quantile-based stratification bins or as a numeric vector providing the bin boundaries (must fully span the range of the stratification variable). For example, for a continuous variable in [0,1], breaks = list(x = 4) indicates stratification into 4 bins, the boundaries of which are determined internally using

quantile(x, probs = seq(0.0, 1.0, length.out = 5))

and breaks = list(x = c(0.0, 0.25, 0.75, 1.0)) specifies a 3 bin structure: [0, 0.25], (0.25, 0.75], and (0.75, 1.0]. Note: the lowest bin is always inclusive.

Examples

# no stratification
no_strat <- create_kfold(mtcars, k = 4L, repeats = 2L)

# stratification on 1 discrete variable
sample_one <- create_kfold(mtcars, k = 4L, repeats = 2L,
                           breaks = list(vs = NULL))

# stratification on 2 variables; 1 continuous + 1 discrete
sample_two <- create_kfold(mtcars, k = 4L, repeats = 2L,
                           breaks = list(gear = 4L, vs = NULL))

# retrieve analysis data for 2nd split
an_2 <- analysis(no_strat, 2L)

# retrieve all splits
an_all <- analysis(no_strat)

# retrieve assessment data for 2nd split
ass_2 <- assessment(no_strat, 2L)

# retrieve all splits
ass_all <- assessment(no_strat)