SIFTED

Select informative features through correlation and genetic search

Syntax

from instancespace.stages.sifted import SiftedStage

outputs = space.run_until_stage(SiftedStage)

Description

SIFTED screens features by their correlation with algorithm performance, clusters candidate features using correlation distance, and searches for representative features with PyGAD. Candidate fitness uses a PILOT projection and KNN cross-validation. It does not use PCA or Random Forest fitness.

The cache belongs to the current genetic-search run. Selected feature indices are mapped back through the correlation-screened columns. Exploration reuses the fitted selection instead of running the search again.

Parameters

sifted.flag enables selection. rho and pval control screening; k, max_iter and replicates control clustering. Genetic-search fields configure population, parents, elitism, crossover, mutation and stopping. The Options Reference documents every field and accepted choice.

The full pipeline derives candidate projection dimensionality from pilot.dims. The direct-call sifted.dims default is 2. A stage seed of None inherits general.seed. n_trees is a compatibility field with no effect on current fitness.

Returns

Selected feature indices and labels, fitted selected data and search diagnostics. See SiftedInput and SiftedOutput.

Examples

from dataclasses import replace
from instancespace.data.options import SiftedOptions

search_options = replace(options, sifted=SiftedOptions.default(
    flag=True, k=3, num_generations=20, sol_per_pop=12,
))

These reduce the example search budget. Increase it only when the data and validation justify more search.