Skip to content

feature-selection.md

Latest commit

 

History

History
49 lines (38 loc) · 1.67 KB

feature-selection.md

File metadata and controls

49 lines (38 loc) · 1.67 KB
Raw

Feature Selection

Language: English
Last updated: 2026-07-12
Switch: Chinese

Overview and Paths

StepwiseSelector and stepwise_selection perform AIC/BIC subset search in forward, backward, or both directions. knockoff_filter and its selector wrappers provide fixed-X and Gaussian second-order model-X FDR control. See the detailed knockoff page.

from statgpu import LinearRegression, StepwiseSelector

selector = StepwiseSelector(
    LinearRegression,
    criterion="bic",
    direction="both",
    max_features=10,
    compute_inference=False,
).fit(X, y)
X_selected = selector.transform(X)

Stepwise Contract

  • Candidate subsets are fitted in sorted feature order, matching predict and transform.
  • Backward selection starts from the full model and enforces max_features as a hard cap before requiring criterion improvement.
  • An intercept-only/null model may win; no feature is forced into the model.
  • Repeated fit() resets histories and caches.
  • n_jobs uses threads so device arrays are not serialized into worker processes.
  • The computation backend and inference capability follow model_class and its keyword arguments.

Outputs

Fitted selectors expose selected_features_, best_model_, aic_history_, bic_history_, and selection_history_, plus predict, transform, and fit_transform.

Validation and Limits

Selection is deterministic for deterministic wrapped estimators. Information-criterion search is combinatorial and is intended for moderate feature counts; knockoff methods are preferable when FDR control or high-dimensional screening is the primary goal.