Skip to main content

tagarnina logo

CI Python License: Apache 2.0

tagarnina is a Python library for feature selection with user-defined constraints. It implements constrained recursive feature elimination (CRFE), where a constraint solver decides which features can be removed at each step, and cross-validated CRFE (CRFECV).

Algorithms

  • CRFE: recursive feature elimination with constraints. Without constraints it behaves like scikit-learn's RFE. Strategies: "probing" (default) and "greedy".
  • CRFECV: CRFE with cross-validation to choose the number of features.

Constraints (in tagarnina.optimization.constraints)

  • MaxCorrelationConstraint: limits the pairwise correlation between selected features.
  • GroupConstraint, AtMostKConstraint, AtLeastKConstraint: bound the number of selected features within a group.
  • MinMaxConstraint: bounds the total number of selected features.
  • OnlyEnforceIfConstraint: if feature1 is selected, feature2 must be selected.
  • OnlyExcludeIfConstraint: if feature1 is selected, feature2 cannot be selected.
  • FixedConstraint: always selects a feature.
  • AllOrNoneConstraint: selects all or none of a set of features.
  • KnapsackConstraint: limits the weighted sum of the selected features.

Table of Contents

Installation

tagarnina requires Python 3.12 or newer. Install the latest release from PyPI:

pip install tagarnina

To install from a local clone of the repository instead:

git clone https://github.com/guillermo-navas-palencia/tagarnina.git
cd tagarnina
pip install .

To install the development tools (tests, linting, type checking):

pip install ".[dev]"

To build the documentation and run the examples (adds Sphinx, matplotlib, and SHAP):

pip install ".[docs]"

Dependencies

Installing the package installs its required dependencies:

Optional extras:

  • dev: pytest, pytest-cov, ruff, mypy
  • docs: sphinx, pydata-sphinx-theme, numpydoc, sphinx-gallery, matplotlib, shap

Getting Started

Example

The example uses the Ames Housing dataset from OpenML, fetched with fetch_openml. It contains 2,930 house sales with 80 features. The target is the continuous sale price, Sale_Price, and there are no missing values. The task selects 20 features under two constraints:

  • AllOrNoneConstraint: one-hot encoded columns of each categorical feature are selected together or not at all.
  • MaxCorrelationConstraint: no two selected numerical features have an absolute Pearson correlation above 0.7.
import pandas as pd
from sklearn.compose import ColumnTransformer, make_column_selector
from sklearn.datasets import fetch_openml
from sklearn.ensemble import RandomForestRegressor
from sklearn.preprocessing import OneHotEncoder

from tagarnina.crfe import CRFE
from tagarnina.optimization.constraints import (
    AllOrNoneConstraint,
    Constraints,
    MaxCorrelationConstraint,
)

X, y = fetch_openml(data_id=43926, as_frame=True, return_X_y=True, parser="liac-arff")

categorical_features = X.select_dtypes("category").columns.tolist()
numerical_features = X.select_dtypes("number").columns.tolist()

print(
    f"{len(categorical_features)} categorical and {len(numerical_features)} numerical features"
)
46 categorical and 34 numerical features

One-hot encode the categorical features:

preprocessor = ColumnTransformer(
    transformers=[
        (
            "cat",
            OneHotEncoder(
                sparse_output=False,
                feature_name_combiner=lambda feature, category: f"{feature}={category}"
         ),
         make_column_selector(dtype_include="category"))
    ],
    remainder="passthrough",
    verbose_feature_names_out=False,
)
X_transformed = pd.DataFrame(
    data=preprocessor.fit_transform(X),
    columns=preprocessor.get_feature_names_out(),
)

print(X_transformed.shape)
(2930, 352)

Create the all-or-none and maximum correlation constraints:

all_or_none_constraints = [
    AllOrNoneConstraint(
        features=[c for c in X_transformed.columns if c.startswith(f"{feature}=")]
    )
    for feature in categorical_features
]

max_corr_constraint = MaxCorrelationConstraint(
    features=numerical_features,
    max_corr=0.7,
    absolute=True,
    method="pearson",
)

constraints = Constraints(
    constraints=[
        max_corr_constraint,
        *all_or_none_constraints
    ]
)

Use random forest regressor as an estimator. Run CRFE with elimination step=0.1 (10% of the features) and R^2 regression score function.

estimator = RandomForestRegressor(max_depth=10, min_samples_leaf=20, random_state=0)

crfe = CRFE(
    estimator=estimator,
    constraints=constraints,
    step=0.1,
    n_features_to_select=20,
    scoring="r2",
    verbose=2,
)
crfe.fit(X_transformed, y)

With verbose=2, the output begins with a banner, the settings, the constraints, and the solver and model details. It then shows a table with one row per iteration and ends with a summary. The last lines are the selected feature names.

Output of the example above (timings vary between machines):

----------------------------------------------------------------------------------------------------------
                                             tagarnina v0.1.0                                             
                                Constrained Recursive Feature Elimination                                 
                          (c) Guillermo Navas-Palencia, 2026, Apache License 2.0                          
----------------------------------------------------------------------------------------------------------
Estimator    : RandomForestRegressor
Data         : 2930 samples, 352 features
Settings     : step=0.1, n_features_to_select=20, strategy=probing, scoring=r2
Constraints  : 47 (MaxCorrelation 1, AllOrNone 46)
Solver       : CP-SAT 9.15.6755, workers=1, time_limit=10s, seed=0
Model        : 704 variables, 628 constraints, built in 0.875s

----------------------------------------------------------------------------------------------------------
 iter   n_features   removed     probes    fit(s)   solve(s)        score   retained       gap  elapsed(s)
----------------------------------------------------------------------------------------------------------
    0          352        78        0|0     4.533      0.041   8.9217e-01     0.9918    0.0082       5.580
    1          274        55        0|0     3.742      0.027   8.9030e-01     1.0000    0.0000       9.434
    2          219        41        0|0     3.324      0.027   8.8920e-01     1.0000    0.0000      12.880
    3          178        57        0|0     3.371      0.021   8.8920e-01     0.9794    0.0206      16.360
    4          121        49        0|0     2.631      0.028   8.8471e-01     0.9739    0.0261      19.126
    5           72        41        0|0     2.651      0.020   8.7967e-01     1.0000    0.0000      21.893
    6           31        11        0|0     1.716      0.014   8.7942e-01     1.0000    0.0000      23.704
    7           20         -          -     1.399          -   8.7879e-01          -         -      25.149
----------------------------------------------------------------------------------------------------------

Termination  : reached the target number of features
Selected     : 20 of 352 features
Probes       : 0 feasible, 0 infeasible, 0 blocked
Time         : wall 25.151s (fit 23.366, solve 0.178, score 0.446), cpu 24.078s
==========================================================================================================

Show the selected features:

selected_features = crfe.get_feature_names_out()
print(selected_features)
['Central_Air=N' 'Central_Air=Y' 'Lot_Area' 'Year_Built' 'Year_Remod_Add'
 'Mas_Vnr_Area' 'BsmtFin_SF_1' 'Bsmt_Unf_SF' 'Total_Bsmt_SF'
 'Second_Flr_SF' 'Gr_Liv_Area' 'Bsmt_Full_Bath' 'Full_Bath'
 'Bedroom_AbvGr' 'Fireplaces' 'Garage_Cars' 'Wood_Deck_SF' 'Open_Porch_SF'
 'Longitude' 'Latitude']

Count the selected categorical and numerical features, including both Central_Air columns.

selected_categorical = []
for feature in categorical_features:
    columns = [c for c in X_transformed.columns if c.startswith(f"{feature}=")]
    n_selected = sum(c in selected_features for c in columns)
    if n_selected:
        selected_categorical.append(feature)
        print(f"{feature}: {n_selected} of {len(columns)} columns selected")

selected_numerical = [c for c in numerical_features if c in selected_features]

print("categorical selected:", len(selected_categorical))
print("numerical selected:", len(selected_numerical))
Central_Air: 2 of 2 columns selected
categorical selected: 1
numerical selected: 18

Citation

If you use tagarnina in academic work, please cite it:

@software{navas_palencia_tagarnina,
  author  = {Navas-Palencia, Guillermo},
  title   = {tagarnina: Python library for feature selection with user-defined constraints},
  year    = {2026},
  url     = {https://github.com/guillermo-navas-palencia/tagarnina},
}

License

Distributed under the Apache License 2.0. See LICENSE for details.

Metadata

Release files for tagarnina 0.1.0

For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.

Source distribution (sdist)

Source distribution for tagarnina 0.1.0
File Size Uploaded Provenance Actions
tagarnina-0.1.0.tar.gz 36.9 kB Without provenance Details for tagarnina-0.1.0.tar.gz

Built distribution (wheel)

Table of built distributions (wheels) for tagarnina 0.1.0
File Interpreter ABI Platform Actions
tagarnina-0.1.0-py3-none-any.whl Python 3 none any Details for tagarnina-0.1.0-py3-none-any.whl

Total release size: 71.1 kB

Release files / tagarnina-0.1.0.tar.gz

Download URL tagarnina-0.1.0.tar.gz
Size 36.9 kB
Tags Source
SHA-256 checksum
How to use checksums
f0cfbabda27bb344eb88a8011121cd90ae241211e61640c131ecd47ea5857332
BLAKE2b-256 checksum
How to use checksums
c21038846a0788b3c5ced35a816b5ed858248372cae9559818de452afaeae201
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.17

Release files / tagarnina-0.1.0-py3-none-any.whl

Download URL tagarnina-0.1.0-py3-none-any.whl
Size 34.2 kB
Tags Python 3
SHA-256 checksum
How to use checksums
2d5be8be4cd8b8b036acbe2852934d42d492abe19293e6641c03a5a8c3edcd7a
BLAKE2b-256 checksum
How to use checksums
7e95698fe75bff2b3f0b56e67c792136c457dd2769a28f36d30c65b1e9763e0b
Upload date
Uploaded using Trusted Publishing?
What is trusted publishing?
No
Uploaded via twine/7.0.0 CPython/3.13.17

Security for tagarnina 0.1.0

Provenance

Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.

No provenance found

None of the files in this release were uploaded with provenance metadata.

A note on package security

PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.

To learn more about protecting yourself from malicious actors, visit our package security documentation.

Release history Release notifications | RSS feed

This release

0.1.0 This release

2 release files No provenance

Anthropic, PBC Visionary sponsor Bloomberg Visionary sponsor Hudson River Trading Visionary sponsor Meta Visionary sponsor NVIDIA Visionary sponsor Microsoft Sustainability sponsor Depot Continuous Integration AWS Cloud computing and Security Sponsor Datadog Monitoring Fastly CDN Google Download Analytics Sentry Error logging StatusPage Status page