moofs

Multi-Objective Optimization for Feature Selection

A Python library for multi-objective feature selection with a unified, scikit-learn compatible API. moofs searches for the best trade-offs between classification error and number of selected features, returns the full Pareto front, and lets you pick the subset that fits your needs.

Installation

pip install moofs

Quick start

from moofs import MOFSSelector

selector = MOFSSelector(algorithm="mofs-rfga", max_evals=5000, random_state=0)
X_reduced = selector.fit_transform(X, y)

selector.pareto_front_   # (n_solutions, 2): [error %, subset size]
selector.support_        # boolean mask of the chosen subset
selector.get_feature_names_out()

It drops into any scikit-learn pipeline:

from sklearn.pipeline import Pipeline
from sklearn.neighbors import KNeighborsClassifier

pipe = Pipeline([
    ("fs", MOFSSelector(max_evals=5000, random_state=0)),
    ("clf", KNeighborsClassifier(n_neighbors=3)),
]).fit(X_train, y_train)

Available algorithms

More algorithms from the MOFS literature (SparseEA, NSGA-II/SDR, SPEA2, MOEA/D, NSPSOFS, CMDPSOFS) are planned for upcoming releases — see the CHANGELOG.

Visualizing Pareto fronts

from moofs import plot_selector, plot_fronts

plot_selector(selector)              # front + highlighted chosen subset
plot_fronts({"MOFS-RFGA": r1, "NSGA-II": r2}, reference=True)

Metrics

Quality indicators follow the PlatEMO definitions used in the MOFS literature, so values are directly comparable with published tables: igd, hv (normalized, reference point (1,1)), coverage (weak dominance), nfs, spacing.

from moofs import compare

table = compare({"MOFS-RFGA": r1, "NSGA-II": r2})
#   algorithm    IGD     HV   NFS  best_error_%  min_subset_size

Research-style API

For experiments and full control over the search:

from moofs import FeatureSelectionProblem, MOFSRFGA, NSGA2

problem = FeatureSelectionProblem(X, y)   # KNN k=3, 3-fold CV, cached
result = MOFSRFGA(problem, pop_size=100, max_evals=300_000, seed=0).run()
result.F        # objective matrix of the Pareto front
result.front    # solutions with binary masks (.x)

The evaluation protocol follows the reference paper: k-NN (k=3) classifier, 3-fold cross-validation, objectives = (classification error %, subset size). Evaluations are memoized; cache hits still count toward max_evals so budgets stay comparable.

Faithfulness notes

Implementations are traceable to their source papers, and ambiguities are documented rather than silently resolved. Notably, the MOFS-RFGA paper's Fig. 1 and its Algorithm 3 disagree on the crossover semantics; moofs defaults to the Fig. 1 reading (consistent with the mutation operator) and exposes interpretation="pseudocode" for the literal alternative. See the documentation for details.

License

MIT — see LICENSE.

Citing

If you use moofs in academic work, please cite the underlying algorithm papers (see the table above). A citable software DOI is planned.