Skip to contents

Buffered Leave-One-Out Cross-Validation

Source: R/borg_spatial_loo.R
borg_spatial_loo.Rd

For each observation, holds it out as the test set and removes all training points within a buffer distance. This ensures spatial independence between train and test in each iteration, following the approach recommended by Roberts et al. (2017).

Usage

borg_spatial_loo(
  data,
  coords,
  buffer = NULL,
  target = NULL,
  diagnosis = NULL,
  max_iter = NULL,
  seed = 42L,
  verbose = FALSE
)

Arguments

data

A data frame.

coords

Character vector of length 2. Coordinate column names.

buffer

Numeric. Exclusion buffer distance (in coordinate units). Training points within this distance of the test point are removed. If NULL, uses the estimated autocorrelation range from borg_diagnose().

target

Character. Response variable name (used for diagnosis if buffer is NULL).

diagnosis

A BorgDiagnosis object. If NULL and buffer is NULL, runs borg_diagnose() to estimate the autocorrelation range.

max_iter

Integer or NULL. Maximum number of LOO iterations. If NULL, uses all n observations. For large datasets, set this to subsample. Default: NULL.

seed

Integer. Random seed for subsampling. Default: 42.

verbose

Logical. Print progress. Default: FALSE.

Value

A list with class "borg_spatial_loo" containing:

folds

List of length n (or max_iter), each with train and test integer index vectors.

buffer

Buffer distance used.

n_excluded

Integer vector: number of training points excluded per iteration.

effective_train_size

Integer vector: training set size per iteration after exclusion.

References

Roberts, D.R., et al. (2017). Cross-validation strategies for data with temporal, spatial, hierarchical, or phylogenetic structure. Ecography, 40(8), 913-929. doi:10.1111/ecog.02881

See also

borg_cv, borg_thin

Examples

set.seed(42)
d <- data.frame(x = runif(50), y = runif(50), z = rnorm(50))
loo <- borg_spatial_loo(d, coords = c("x", "y"), buffer = 0.2)
length(loo$folds)  # 50 iterations
mean(loo$n_excluded)  # avg points excluded per iteration