Generate Cross-validation Folds

Creates cross-validation folds by splitting data based on specified design variables or stratification factors. This function helps in creating stratified cross-validation splits while respecting the structure of the data, particularly for complex experimental designs.

Usage

fold_over(x, ...)

# S3 method for class 'multidesign'
fold_over(x, ..., preserve_row_ids = FALSE)

# S3 method for class 'multiframe'
fold_over(x, ..., preserve_row_ids = FALSE)

Arguments

x

The dataset to fold over (multidesign, hyperdesign, matrix, etc.)

...

Additional arguments passed to methods: * For multidesign/hyperdesign: unquoted names of variables to split on * For multidesign/hyperdesign/multiframe: `preserve_row_ids = TRUE` to carry original source row ids into fold outputs via a reserved `.orig_index` column * For matrices: nfolds (number of folds), stratum (stratification variable)

preserve_row_ids

Logical; if `TRUE`, carry original source row ids into fold `analysis` and `assessment` designs via a reserved `.orig_index` column. Matching `held_out$row_ids` metadata is also included.

Value

A foldlist object containing:

analysis

Training data for each fold

assessment

Test data for each fold

held_out

Information about which values were held out in each fold

Details

The function creates folds by splitting the data based on unique combinations of the specified variables. For each fold, one combination is held out as the assessment set (test set), while the rest form the analysis set (training set).

Different behaviors are implemented for different object types: * For multidesign objects: Creates folds based on design variables * For hyperdesign objects: Creates folds across blocks, respecting block structure * For matrices: Creates folds based on external stratification variables

See also

multidesign for creating multidesign objects, hyperdesign for creating hyperdesign objects, split_indices for splitting indices without creating folds

Examples

# With a multidesign object
X <- matrix(rnorm(100*10), 100, 10)
groups <- rep(1:2, each=50)
mds <- multidesign(X, data.frame(group=groups, subject=rep(1:10, each=10)))
folds_by_group <- fold_over(mds, group)

# With a hyperdesign object (multiple subjects)
d1 <- multidesign(matrix(rnorm(10*20), 10, 20),
                 data.frame(condition=rep(c("A","B"), 5), run=1:10))
d2 <- multidesign(matrix(rnorm(10*20), 10, 20),
                 data.frame(condition=rep(c("A","B"), 5), run=1:10))
hd <- hyperdesign(list(d1, d2), block_names=c("subject1", "subject2"))
folds_by_condition <- fold_over(hd, condition)