Feature Selection

fyron.feature_selection provides transparent feature-selection strategies for clinical data science. It is designed for model development notebooks where every selection decision should be inspectable and reportable.

Install

bash

uv add "fyron[feature-selection]"

Imports

python

from fyron import feature_selection as fs

Function Reference

Function	Required	Optional	Returns
`univariate_feature_screening(X, y, ...)`	feature table, target	task, alpha, FDR correction	selected features and screening table
`variance_filter(X, ...)`	feature table	variance threshold	selected features and variance table
`correlation_filter(X, ...)`	feature table	threshold, method	selected/dropped features and correlation pairs
`model_importance_selection(estimator, X, y, ...)`	estimator, features, target	threshold or top N	importance table and fitted model
`permutation_importance_selection(estimator, X, y, ...)`	estimator, features, target	repeats, scoring, threshold/top N	permutation importance table
`boruta_feature_selection(X, y, ...)`	features, target	Boruta/sklearn options	Boruta ranking table
`stability_selection(estimator, X, y, ...)`	estimator, features, target	rounds, sample fraction, threshold	selection-frequency table
`combine_feature_selection_results(results, ...)`	result dictionaries	strategy, minimum methods	consensus table

Example

python

from sklearn.ensemble import RandomForestClassifier
from fyron import feature_selection as fs

univariate = fs.univariate_feature_screening(
    X,
    y,
    task="classification",
    alpha=0.05,
)

correlation = fs.correlation_filter(
    X,
    threshold=0.85,
    method="spearman",
)

stable = fs.stability_selection(
    RandomForestClassifier(n_estimators=300, random_state=42),
    X[correlation["selected_features"]],
    y,
    n_rounds=100,
    selection_threshold=0.6,
)

final = fs.combine_feature_selection_results(
    {
        "univariate": univariate,
        "correlation": correlation,
        "stability": stable,
    },
    strategy="consensus",
    min_methods=2,
)

final["selected_features"]
final["summary_table"]

Interpretation

Feature selection is a model-development tool, not causal evidence. Use the returned tables to document why a feature was kept or dropped, and validate the final feature set on held-out or external data when possible.