Specify a reproducible SOM ensemble

Usage

som_spec(
  grids = data.frame(xdim = 7L, ydim = 5L),
  seeds = 1:10,
  rlen = 500L,
  alpha = c(0.05, 0.01),
  radius = NULL,
  topology = c("hexagonal", "rectangular"),
  neighbourhood = c("gaussian", "bubble"),
  toroidal = FALSE,
  mode = c("batch", "online", "pbatch"),
  k = 2:8,
  max_na_fraction = 0,
  layer_weights = NULL,
  normalize_layers = TRUE,
  cores = 1L
)

Arguments

grids: A two-column data frame with xdim and ydim, a numeric vector of length two for one grid, or a list of such vectors.
seeds: Integer random seeds.
rlen: Number of training iterations.
alpha: Initial and final learning rates.
radius: Optional initial and final neighbourhood radii.
topology: Rectangular or hexagonal grid.
neighbourhood: Bubble or Gaussian neighbourhood.
toroidal: Whether the grid is toroidal.
mode: Training mode supported by kohonen::supersom().
k: Candidate numbers of hard partitions retained for later auditing.
max_na_fraction: Maximum allowed fraction of missing variables per layer and sample passed to kohonen.
layer_weights: Optional non-negative scientific weights, one per layer.
normalize_layers: Whether to divide each scientific layer weight by its training-set mean squared Euclidean distance. The package uses the variance identity for a deterministic, leakage-safe estimate and applies the same effective weights to SOM and cross-model fits.
cores: Number of cores used by kohonen in parallel batch mode.

Value

A som_spec object.

Examples

specification <- som_spec(
  grids = list(c(4, 3), c(5, 4)), seeds = 1:2, rlen = 50, k = 2:4
)
expand_som_spec(specification)
#>   xdim ydim grid_id seed   model_id
#> 1    4    3       1    1 g01_4x3_s1
#> 2    4    3       1    2 g01_4x3_s2
#> 3    5    4       2    1 g02_5x4_s1
#> 4    5    4       2    2 g02_5x4_s2