evo-imp - Evolutionary Symbolic Imputer

A scikit-learn-compatible missing-value imputer for tabular data, powered by DEAP. evo-imp evolves an interpretable symbolic expression per feature with Genetic Programming, capturing non-linear relationships that SimpleImputer/KNNImputer/IterativeImputer cannot represent. It is a natural first step in the pipeline, upstream of evo-gpfe (feature engineering) and evo-gafs (feature selection).

Part of the evo-suite family (import name: evo_imp). Documentation: https://evo-suite.readthedocs.io/

Why evo-imp?

Installation

pip install evo-imp            # core
pip install "evo-imp[viz]"     # + matplotlib for the plotting helpers
pip install "evo-imp[simplify]"  # + sympy for simplify_expressions=True

Quickstart

from sklearn.datasets import load_diabetes
from evo_imp import EvoImpConfig, EvoImputer, introduce_mcar

X = load_diabetes(as_frame=True).data
X_missing, _ = introduce_mcar(X.values, missing_rate=0.20, random_seed=42)

imputer = EvoImputer(
    config=EvoImpConfig(population_size=150, n_generations=30, verbose=False)
)
X_imputed = imputer.fit_transform(X_missing)

print(imputer.summary())
print(imputer.get_expressions())

Unlike the rest of evo-suite, EvoImputer is unsupervised (fit(X, y=None)) and its whole purpose is to accept NaN input — see the Concepts guide for why this is a deliberate exception to the family's usual scikit-learn conventions.

Documentation & examples

• Full documentation (user guide + API reference): https://evo-suite.readthedocs.io/
• Runnable examples: the repository's examples/evo-imp/ directory.

Citation

@software{evo_imp,
  author    = {Skrauba, Axel},
  title     = {evo-imp: Evolutionary Symbolic Imputer for tabular data},
  year      = {2026},
  version   = {0.1.0},
  url       = {https://github.com/AxelSkrauba/evo-suite}
}

License

MIT