Split Data into Approximately Balanced Folds — make

Create a fold-assignment vector for random or spatially structured cross-validation.

Usage

make_blocks(
  nfolds = 5,
  df = data.frame(),
  nblocks = nfolds * 2,
  npoints = NA,
  pres = numeric()
)

Arguments

nfolds: Numeric. Number of folds to create.
df: Optional data.frame. Columns define the structure used to build folds. When omitted, random folds are created from npoints.
nblocks: Numeric. Number of initial clusters used to build the folds. Must be at least nfolds.
npoints: Optional numeric. Number of observations to split when df is not supplied.
pres: Optional binary vector. Used to restrict CLARA clustering to a subset of rows and assign the remainder by nearest neighbours.

Value

An integer vector of length nrow(df) or npoints, giving the fold assignment for each observation.

Details

If df has one column, folds are based on quantile bins. If df has two or more columns, folds are based on CLARA clustering. Remaining clusters are assigned to folds with an optimization step that tries to balance fold sizes as evenly as possible.

References

Brun, P., Thuiller, W., Chauvier, Y., Pellissier, L., Wüest, R. O., Wang, Z., & Zimmermann, N. E. (2020). Model complexity affects species distribution projections under climate change. Journal of Biogeography, 47(1), 130-142.

Examples

if (FALSE) { # \dontrun{
# Downloading worldwide the observations of Panthera tigris
obs.pt <- get_gbif(
    sp_name = "Panthera tigris",
    basis = c("OBSERVATION","HUMAN_OBSERVATION",
        "MACHINE_OBSERVATION","OCCURRENCE")
)

# Create a vector of folds (n = 5) spatially blocked (n = 10)
block.pt <- make_blocks(
    nfolds = 5,
    df = obs.pt[, c("decimalLatitude","decimalLongitude")],
    nblocks = 10
)

# Plot one colour per fold
countries <- rnaturalearth::ne_countries(
    type = "countries",
    returnclass = "sv"
)
countries.focus <- terra::crop(
    countries,
    terra::ext(60,100,0,40)
)
terra::plot(countries.focus, col = "#bcbddc")
graphics::points(
    obs.pt[, c("decimalLongitude","decimalLatitude")],
    pch = 20,
    col = block.pt,
    cex = 1
)
} # }