MLX-backed elastic net regression — mlxs

Fit lasso or elastic-net penalised regression paths using MLX arrays for the heavy linear algebra. Dense Gaussian and binomial paths stay on the MLX backend throughout the iterative updates, with repeated chunk updates traced through Rmlx::mlx_compile() to reduce host overhead.

Usage

mlxs_glmnet(
  x,
  y,
  family = mlxs_gaussian(),
  alpha = 1,
  lambda = NULL,
  nlambda = 100,
  lambda_min_ratio = 1e-04,
  standardize = TRUE,
  intercept = TRUE,
  use_strong_rules = TRUE,
  maxit = 1000,
  tol = 1e-06
)

Arguments

x: Numeric matrix of predictors (observations in rows).
y: Numeric response vector (for binomial, values must be 0/1).
family: MLX-aware family object, e.g. mlxs_gaussian() or mlxs_binomial().
alpha: Elastic-net mixing parameter (1 = lasso, currently alpha must be strictly positive).
lambda: Optional decreasing sequence of penalty values. If NULL, a sequence of length nlambda is generated from lambda_max down to lambda_max * lambda_min_ratio.
nlambda: Length of automatically generated lambda path.
lambda_min_ratio: Smallest lambda as a fraction of lambda_max.
standardize: Should columns of x be scaled before fitting?
intercept: Should an intercept be fit?
use_strong_rules: Retained for API compatibility. The dense MLX solver keeps all coefficients on device, so this flag currently does not change the computation.
maxit: Maximum proximal-gradient iterations per lambda value.
tol: Convergence tolerance on the coefficient updates.

Value

An object of class mlxs_glmnet containing the fitted coefficient path, intercepts, lambda sequence, and scaling information.

Note

This implementation uses dense MLX updates rather than glmnet's coordinate-descent Fortran kernels, so glmnet::glmnet() can still be faster on smaller problems. For very tall Gaussian problems, the fitter switches to a covariance-space MLX solver to reduce repeated X'X products along the path.