felimination 0.8.0

This library contains some useful scikit-learn compatible classes for feature selection.

Felimination

This library contains some useful scikit-learn compatible classes for feature selection.

Features

• Recursive Feature Elimination with Cross Validation using Permutation Importance
• Forward Feature Selection with Minimum Redundancy Maximum Relevance (MRMR)
• Hybrid Genetic Algorithms x Feature Importance selection

Requirements

• Python 3.10+
• NumPy
• Scikit-learn
• Pandas

Installation

In a terminal shell run the following command

pip install felimination

Usage

Recursive Feature Elimination

In this section it will be illustrated how to use the PermutationImportanceRFECV class.

import numpy as np
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold, cross_validate

from felimination.callbacks import plot_progress_callback
from felimination.rfe import PermutationImportanceRFECV


X, y = make_classification(
    n_samples=1000,
    n_features=200,
    n_informative=6,
    n_redundant=10,
    n_clusters_per_class=1,
    random_state=42,
    shuffle=False
)

model = LogisticRegression(random_state=42)

selector = PermutationImportanceRFECV(
    model,
    step=0.2,
    callbacks=[plot_progress_callback],
    scoring="roc_auc",
    cv=StratifiedKFold(random_state=42, shuffle=True),
    best_iteration_selection_criteria="mean_test_score"
)

selector.fit(X, y)

selector.support_
# array([False,  True,  True,  True, False,  True,  True,  True,  True,
#         True,  True, False, False, False, False,  True, False,  True,
#        False,  True, False, False, False, False, False, False, False,
#         True, False, False, False, False, False, False, False, False,
#         True, False, False, False, False, False,  True, False, False,
#         True, False,  True, False, False, False, False, False, False,
#        False, False, False, False, False, False, False, False, False,
#        False, False,  True,  True, False, False,  True, False,  True,
#        False,  True, False, False,  True, False, False, False,  True,
#        False,  True, False, False, False,  True, False,  True, False,
#        False, False, False,  True, False,  True, False, False, False,
#        False, False, False,  True, False, False,  True,  True,  True,
#        False, False, False, False, False,  True, False, False, False,
#        False, False, False, False, False, False, False, False, False,
#        False, False, False, False, False, False, False, False, False,
#        False, False, False, False, False, False, False, False, False,
#        False,  True, False, False, False, False, False, False, False,
#        False, False,  True,  True, False, False, False,  True, False,
#        False, False, False, False, False, False,  True, False, False,
#        False, False, False, False, False,  True, False, False, False,
#         True, False,  True, False, False, False,  True, False, False,
#        False, False, False, False, False, False, False, False,  True,
#        False, False])

selector.ranking_
# array([23,  5, 12,  2, 24, 15, 16,  1,  3,  6,  4, 24, 23, 19, 24, 14, 19,
#        17, 21, 16, 24, 20, 24, 21, 24, 18, 22, 16, 22, 23, 24, 21, 22, 22,
#        21, 22, 16, 20, 23, 23, 24, 20, 13, 24, 23, 13, 23, 14, 23, 22, 22,
#        24, 19, 19, 23, 19, 23, 20, 23, 23, 22, 23, 23, 23, 24, 17, 11, 20,
#        23, 10, 22, 14, 18, 13, 24, 21, 12, 23, 24, 18,  9, 21, 13, 21, 24,
#        21, 16, 18, 15, 21, 24, 22, 20, 17, 20, 17, 22, 21, 24, 19, 19, 24,
#        16, 20, 24, 15, 17, 17, 24, 24, 24, 22, 21, 14, 21, 22, 23, 24, 21,
#        21, 22, 20, 23, 23, 24, 20, 23, 23, 24, 24, 18, 19, 20, 22, 23, 24,
#        22, 18, 21, 24, 24, 23, 22, 24, 22, 15, 20, 21, 23, 23, 22, 19, 22,
#        20, 22,  8, 12, 20, 23, 22, 17, 18, 23, 24, 24, 22, 21, 24, 11, 19,
#        20, 24, 21, 24, 18, 21, 16, 21, 19, 24, 17, 18, 15, 24, 22, 24, 10,
#        19, 22, 24, 23, 24, 23, 20, 24, 23, 19,  7, 18, 23])
selector.plot()

Forward Feature Selection with MRMR

In this section it will be illustrated how to use the MRMRCV class, which performs forward feature selection using the Minimum Redundancy Maximum Relevance (MRMR) criterion.

MRMR scores candidate features by combining two quantities:

• Relevance: how much information a feature shares with the target (mutual information by default).
• Redundancy: how much information a feature shares with already-selected features.

This actively avoids picking correlated features: once a relevant feature is selected, correlated copies are penalised and deprioritised.

Redundancy aggregation

When multiple features have already been selected, each candidate's redundancy score is computed against each of them individually and then aggregated into a single value. The redundancy_aggregation parameter controls how:

Value	Behaviour
'max' (default)	Element-wise maximum — a candidate is penalised as soon as it is highly redundant with any selected feature.
'mean'	Element-wise mean — matches the formulation in the original MRMR paper (Peng et al., 2005).
callable	A custom function f(matrix) -> array where matrix has shape (n_selected, n_features).

Note: The default 'max' deviates from the original MRMR paper, which uses the mean. 'max' is chosen as the default because it more aggressively penalises features that duplicate information already captured by any single selected feature, which tends to work better in practice for forward selection with CV scoring.

from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import StratifiedKFold

from felimination.callbacks import plot_progress_callback
from felimination.mrmr import MRMRCV


X, y = make_classification(
    n_samples=1000,
    n_features=200,
    n_informative=6,
    n_redundant=10,
    n_clusters_per_class=1,
    random_state=42,
    shuffle=False
)

model = LogisticRegression(random_state=42)

selector = MRMRCV(
    model,
    step=0.2,
    max_features_to_select=50,
    callbacks=[plot_progress_callback],
    scoring="roc_auc",
    cv=StratifiedKFold(random_state=42, shuffle=True),
    best_iteration_selection_criteria="mean_test_score",
    random_state=42,
    min_relevance=0.1,
    redundancy_aggregation="max",  # default; use 'mean' for original MRMR behaviour

)

selector.fit(X, y)

selector.support_
# array([False, False,  True, False,  True, False, False, False,  True,
#        False,  True, False,  True, False,  True, False, False, False, ...])

selector.plot()

Genetic Algorithms

In this section it will be illustrated how to use the HybridImportanceGACVFeatureSelector class.

import numpy as np
from sklearn.datasets import make_classification
from sklearn.linear_model import LogisticRegression

from felimination.ga import HybridImportanceGACVFeatureSelector


# Create dummy dataset
X, y = make_classification(
    n_samples=1000,
    n_features=20,
    n_informative=6,
    n_redundant=10,
    n_clusters_per_class=1,
    random_state=42,
)

# Initialize selector
selector = HybridImportanceGACVFeatureSelector(
    LogisticRegression(random_state=42),
    random_state=42,
    pool_size=5,
    patience=5
)

# Run optimisation
selector.fit(X, y)

# Show selected features
selector.support_
#array([False,  True, False,  True,  True, False, False, False,  True,
#       False, False, False,  True,  True,  True,  True, False,  True,
#        True, False])

# Show best solution
selector.best_solution_
# {'features': [1, 12, 13, 8, 17, 15, 18, 4, 3, 14],
#  'train_scores_per_fold': [0.88625, 0.89, 0.8825, 0.8925, 0.88625],
#  'test_scores_per_fold': [0.895, 0.885, 0.885, 0.89, 0.89],
#  'cv_importances': [array([[ 1.09135972,  1.13502636,  1.12100231,  0.38285736,  0.28944072,
#            0.04688614,  0.44259813,  0.09832365,  0.10190421, -0.48101593]]),
#   array([[ 1.17345812,  1.29375208,  1.2065342 ,  0.40418709,  0.41839714,
#            0.00447802,  0.466717  ,  0.21733829, -0.00842075, -0.50078996]]),
#   array([[ 1.15416104,  1.18458564,  1.18083266,  0.37071253,  0.22842685,
#            0.1087814 ,  0.44446793,  0.12740545,  0.00621562, -0.54064287]]),
#   array([[ 1.26011643,  1.36996058,  1.30481424,  0.48183549,  0.40589887,
#           -0.01849671,  0.45606913,  0.18330816,  0.03667055, -0.50869557]]),
#   array([[ 1.18227123,  1.28988253,  1.2496398 ,  0.50754295,  0.38942303,
#           -0.01725074,  0.4481891 ,  0.19472963,  0.10034316, -0.50131192]])],
#  'mean_train_score': 0.8875,
#  'mean_test_score': 0.889,
#  'mean_cv_importances': array([ 1.17227331,  1.25464144,  1.21256464,  0.42942709,  0.34631732,
#          0.02487962,  0.45160826,  0.16422104,  0.04734256, -0.50649125])}

# Show progress as a plot
selector.plot()

Looks like that the optimisation process converged after 2 steps, since the best score did not improve for 5(=patience) consecutive steps, the optimisation process stopped early.

License

This project is licensed under the BSD 3-Clause License - see the LICENSE.md file for details

Acknowledgments

• scikit-learn