Multi-objective selection (NSGA-II)

Single-objective mode collapses performance and compression into one number via alpha. Multi-objective mode keeps them separate and optimises both at once with NSGA-II, returning the full Pareto front of non-dominated solutions.

from sklearn.datasets import load_breast_cancer
from sklearn.tree import DecisionTreeClassifier
from evo_gafs import GAConfig, GAFeatureSelector

X, y = load_breast_cancer(return_X_y=True, as_frame=True)

selector = GAFeatureSelector(
    estimator=DecisionTreeClassifier(random_state=42),
    config=GAConfig(mode="multiobjective", population_size=40, n_generations=30, verbose=False),
)
selector.fit(X, y)

for point in sorted(selector.result_.pareto_front, key=lambda p: p["n_features"]):
    print(point["n_features"], round(point["cv_score"], 4))

Each entry of result_.pareto_front is a dict with mask, cv_score, compression and n_features. Because the front contains non-dominated trade-offs, you can choose the smallest subset that meets your accuracy target, or the most accurate subset under a size budget.

Visualising the front

With the viz extra installed:

from evo_gafs import GAPlotter

fig = GAPlotter.plot_pareto_front(selector.result_)
fig.savefig("pareto.png", dpi=120, bbox_inches="tight")

When to use which mode

• Single-objective when you already know the trade-off you want (set alpha).

• Multi-objective when you want to discover the trade-off curve and decide afterwards, or when reporting the full performance/size frontier matters (e.g. in research).