Grouped Stability Selection for Functional Predictors — stability_selection

Repeatedly subsamples observations, refits a sparse base selector, and computes exact feature- and group-level selection frequencies. This is the generic FDA recipe for basis expansions, discretized curves, or FPCA scores.

Usage

stability_selection_fda(
  x,
  y = NULL,
  selector = "group_lasso",
  selector_fun = NULL,
  groups = NULL,
  family = c("gaussian", "binomial"),
  B = 100L,
  sample_fraction = 0.5,
  cutoff = 0.75,
  seed = NULL,
  keep_subsamples = FALSE,
  ...
)

Arguments

x: Any input accepted by as_functional_matrix(), or an fda_design object.
y: Response vector. Leave as NULL when x is an fda_design.
selector: Either "lasso", "group_lasso", "sparse_group_lasso", one of the backend-specific aliases ("glmnet", "grpreg", "sgl"), or a custom function.
selector_fun: Optional custom selector. It must accept X, y, groups, and family, and return either a coefficient vector or a logical selection vector of length p.
groups: Optional grouping structure. Defaults to block-level groups when x is supplied as a list, and otherwise to one group per feature.
family: Model family passed to the built-in selectors.
B: Number of subsampling replicates.
sample_fraction: Fraction of observations drawn without replacement in each subsample.
cutoff: Stability threshold used to define selected_features and selected_groups.
seed: Optional random seed.
keep_subsamples: Should the sampled row indices be returned?
...: Additional arguments forwarded to the built-in or custom selector.

Value

An object of class fda_stability_selection.