Penalized linear discriminant analysis

Penalized Fisher's linear discriminant (Witten & Tibshirani 2011) with L1 or fused-lasso penalties, multi-class support, and built-in cross-validation.

Usage

plda(x, ...)

# Default S3 method
plda(
  x,
  y,
  K = NULL,
  lambda = NULL,
  penalty = c("L1", "fused"),
  lambda2 = NULL,
  autotune = TRUE,
  nfold = 5L,
  lambda_grid = NULL,
  maxit = 100L,
  tol = 0.000001,
  na.action = c("impute", "omit"),
  encoding = c("onehot", "means", "integer", "auto"),
  encoding_spec = NULL,
  nthreads = 0L,
  ...
)

# S3 method for class 'formula'
plda(formula, data = NULL, ...)

Arguments

x: Numeric predictor matrix or a data frame (factor / character columns are expanded to a numeric model matrix via model.matrix(~ ., x), with the intercept column dropped), or a formula for the formula interface.
...: Passed to methods.
y: Factor (or coercible) class label of length nrow(x).
K: Number of discriminant vectors (<= G-1). If supplied, K is fixed; if NULL (default), K is chosen by cross-validation along with lambda. When autotune = FALSE, NULL means G-1.
lambda: Magnitude penalty. Required when autotune = FALSE.
penalty: "L1" (default) or "fused".
lambda2: Fused-lasso difference penalty (used when penalty = "fused").
autotune: If TRUE (default), cross-validate lambda (and K).
nfold: CV folds.
lambda_grid: Optional CV grid.
maxit: MM solver maximum iterations.
tol: MM solver convergence tolerance.
na.action: Strategy for missing values in the predictors x. "impute" (default) replaces each NA with that column's median (the medians are stored on the fit and reapplied at predict()); "omit" drops rows with any NA in x (subsetting y to match). Either way a warning is emitted. NaN / Inf in x are always a hard error, and a missing class label y always hard-stops (y is never imputed or dropped). For the formula method the model frame is built with na.action = na.pass and this rule is applied downstream. When x carries factor / character columns, NAs in those columns are imputed on the expanded indicator scale (median of the 0/1 dummies), not on the original level; pass na.action = "omit" to drop such rows if that matters.
encoding: How factor / character columns in a data-frame x are converted to numeric predictors: "onehot" (default) reproduces the legacy model.matrix(~ ., x) indicator expansion exactly; "means", "integer", and "auto" defer to rr_encode_factors(). The recipe is stored on the fit and replayed at predict(). Ignored for a numeric matrix x or the formula interface.
encoding_spec: Optional explicit encoding recipe passed to rr_encode_factors() as spec; when supplied the default "onehot" model-matrix path is bypassed.
nthreads: Integer. Number of worker threads for the cross-validation autotune (autotune = TRUE). 0 (the default, taken from getOption("roadrunner.nthreads") when set) uses RcppParallel::defaultNumThreads(). The CV fold loop is parallelised with fixed per-fold output slots and a serial-order reduction, so the fitted discriminants are byte-identical regardless of nthreads. Ignored when autotune = FALSE (a single fit is already sequential).
formula: A model formula; response on the left, predictors on the right.
data: A data frame (or environment) containing the formula variables.

Value

An object of S3 class "plda".