Fit Sinkhorn OT cohort barycenter scorer.

Usage

fit_sinkhorn_ot_scorer(
  X_train,
  y_train = NULL,
  meta_train,
  cohort_col = "cohort",
  panel_size = 20L,
  panel_features = NULL,
  feature_weights = NULL,
  epsilon = NULL,
  lambda = NULL,
  epsilon_grid = c(0.05, 0.1, 0.2),
  lambda_grid = c(0.35, 0.6, 0.85),
  tune = TRUE,
  tune_max_folds = 3L,
  tune_n_atoms = 120L,
  tune_max_cohort_atoms = 60L,
  n_barycenter_atoms = 1000L,
  max_cohort_atoms = 250L,
  barycenter_iter = 4L,
  max_sinkhorn_iter = 250L,
  sinkhorn_tol = 1e-07,
  min_pivot_features = 12L,
  max_pivot_features = 80L,
  pseudocount = NULL,
  seed = 42L,
  backend = c("auto", "python", "R")
)

Arguments

X_train: Samples x features non-negative matrix.
y_train: Optional 0/1 outcome vector. Required when panel selection or hyperparameter tuning is requested.
meta_train: data.frame with one row per sample and a cohort column.
cohort_col: Column identifying training cohorts.
panel_size: Number of features selected when `panel_features` is NULL.
panel_features: Optional fixed panel features.
feature_weights: Optional named numeric panel weights.
epsilon: Entropic-transport regularization. If NULL, selected from `epsilon_grid` when tuning is enabled.
lambda: Barycenter shrinkage weight. If NULL, selected from `lambda_grid` when tuning is enabled.
epsilon_grid, lambda_grid: Candidate hyperparameter grids.
tune: Logical. If TRUE, use train-only nested cohort CV to tune transport hyperparameters.
tune_max_folds: Maximum training folds used during tuning.
tune_n_atoms: Number of barycenter atoms used during tuning.
tune_max_cohort_atoms: Maximum cohort atoms used during tuning.
n_barycenter_atoms: Number of free-support barycenter atoms.
max_cohort_atoms: Maximum samples retained per cohort cloud.
barycenter_iter: Number of barycenter updates.
max_sinkhorn_iter: Maximum Sinkhorn iterations.
sinkhorn_tol: Sinkhorn convergence tolerance.
min_pivot_features, max_pivot_features: Bounds for CLR pivot features.
pseudocount: Optional additive pseudocount before CLR transform.
seed: Integer random seed.
backend: Backend for barycenter/projection computation.

Value

A frozen `sinkhorn_ot_scorer` object.