Tucker Decomposition for Multi-Dimensional Data

step_measure_tucker() creates a specification of a recipe step that applies Tucker decomposition to multi-dimensional measurement data, extracting component scores as features for modeling.

Usage

step_measure_tucker(
  recipe,
  ...,
  ranks = 3L,
  center = TRUE,
  scale = FALSE,
  max_iter = 500L,
  tol = 1e-06,
  prefix = "tucker_",
  role = "predictor",
  trained = FALSE,
  skip = FALSE,
  id = recipes::rand_id("measure_tucker")
)

Arguments

recipe: A recipe object.
...: One or more selector functions to choose measure columns. If empty, all nD measure columns are used.
ranks: A vector of ranks for each mode. If a single integer, the same rank is used for all modes. Default is 3.
center: Logical. Should data be centered before decomposition? Default is TRUE.
scale: Logical. Should data be scaled before decomposition? Default is FALSE.
max_iter: Maximum number of iterations. Default is 500.
tol: Convergence tolerance. Default is 1e-6.
prefix: Prefix for output column names. Default is "tucker_".
role: Not used.
trained: Logical indicating if the step has been trained.
skip: Logical. Should the step be skipped when baking?
id: Unique step identifier.

Value

An updated recipe with the new step added.

Details

Tucker decomposition (also known as higher-order SVD or multilinear SVD) decomposes a tensor into a core tensor multiplied by factor matrices along each mode. Unlike PARAFAC, Tucker allows different ranks for each mode, providing more flexibility.

Requirements

Input must be measure_nd_list with 2+ dimensions
All samples must have the same grid (regular, aligned)
The multiway package must be installed (in Suggests)

Output

Creates numeric feature columns representing the unfolded core tensor scores for each sample.

Note

This step requires the multiway package. Install with: install.packages("multiway")

Examples

if (FALSE) { # \dontrun{
library(recipes)

# After ingesting 2D data as nD measurements
rec <- recipe(concentration ~ ., data = lc_dad_data) |>
  step_measure_input_long(
    absorbance,
    location = vars(time, wavelength)
  ) |>
  step_measure_tucker(ranks = c(5, 3)) |>
  prep()

bake(rec, new_data = NULL)
} # }