uefds 0.1.0

Universal Evolutionary Feature Discovery and Selection Framework

UEFDS — Universal Evolutionary Feature Discovery and Selection Framework

A domain-agnostic, two-stage evolutionary system for automatic feature engineering: expression-tree genetic programming discovers candidate features (Stage 1), then a genetic-algorithm wrapper selects the best subset for a real model (Stage 2, in progress).

This package currently implements Stage 1.

Install

pip install -e .

(editable install — recommended while you're actively developing this)

For development/testing extras:

pip install -e ".[dev]"

Quick start

from sklearn.datasets import load_breast_cancer
from uefds import Stage1Config, run_stage1

data = load_breast_cancer()
X, y = data.data, data.target
feature_names = list(data.feature_names)

config = Stage1Config(
    population_size=50,
    n_generations=30,
    task="classification",       # or "regression"
    protected_features=[0],      # optional: guarantee feature 0 never disappears
)

result = run_stage1(X, y, feature_names, config)

print(result["hall_of_fame"].summary(n=10))

Run the full example:

python examples/demo_stage1.py

Package layout

src/uefds/
├── operators.py          operator library + operator-learning tracker
├── tree.py                 expression tree chromosome + breeding operators
├── fitness.py               Quality / Diversity / Simplicity scoring
├── hall_of_fame.py         archive + lineage tracking
└── stage1_discovery.py     main generation loop

Status

• Stage 1 — Evolutionary Feature Discovery
• Compression layer (correlation/MI clustering dedup)
• Stage 2 — Evolutionary Feature Selection (GA wrapper + XGBoost + SHAP)
• NSGA-II multi-objective selection mode
• Validation framework (CV / time-based / out-of-time splits)

Requirements

• Python >= 3.10
• numpy >= 1.24
• scikit-learn >= 1.3