Bidirectional Floating Feature Selection Guided by Uncertainty Quantification

Authors: Marcos López-De-Castro, José González-Gomariz, Alberto García-Galindo Farnoosh Abbas-Aghababazadeh, Kewei Ni, Benjamin Haibe-Kains, Ruben Armañanzas Arnedillo,

Contact: mlopezdecas@unav.es

Description: Conformal prediction–driven feature selection, with applications in immuno-oncology datasets. We propose a novel bidirectional floating algorithm for feature selection named Conformal Bidirectional Floating Search Algorithm (CBFS), in which the feature search is enhanced by information from the conformal prediction framework.

Set Up

Prerequisites

Clone environment in pixi.toml with

pixi install

or install via:

pip install cbfs

Usage

from cbfs import ffs


data_path = "your_path.h5ad" # .csv are also supported
target_column =  "target-column-name"
run_id= 0 # equivalent to random seed (int)

ffs_instance = ffs.FloatingFeatureSelector(run_id=run_id, data_path=data_path, target_column=target_column)
experiment_result = ffs_instance.run_ffs(n_feat=10)


print("Experiment completed successfully!")

print("Selected features:", experiment_result)

coverage = ffs_instance.Empirical_coverage_
uncertainty = ffs_instance.Uncertainty_
certainty = ffs_instance.Certainty_

print(f"Empirical coverage: {coverage}")
print(f"Uncertainty: {uncertainty}")
print(f"Certainty: {certainty}")

all_results[run_id] = {"selected_features": experiment_result, "run_id": run_id, "empirical_coverage": coverage,
                    "uncertainty": uncertainty, "certainty": certainty}