Skip to content

Feature Selection via Genetic Algorithm (FSGA)

A university project implementing feature selection using Genetic Algorithms, with evaluation and visualization tools.

Tests PyPI version Total Downloads PyPI - Downloads Python 3.11+ License: MIT

Overview

FSGA is a modular library for feature selection using genetic algorithms.

Key Features

  • Modular Design: Swappable operators, selectors, and evaluators
  • Multiple Operators: 4 crossover operators, 4 selection strategies, 3 fitness functions
  • Baseline Comparisons: Built-in RFE, LASSO, Mutual Information, Chi², ANOVA
  • Statistical Testing: Wilcoxon, Mann-Whitney, Cohen's d, Jaccard stability
  • Visualization: 9 plot functions for analysis and comparison
  • Experiment Framework: ExperimentRunner for reproducible experiments
  • Configuration: YAML-based configuration system

Quick Example

from fsga.core.genetic_algorithm import GeneticAlgorithm
from fsga.datasets.loader import load_dataset
from fsga.evaluators.accuracy_evaluator import AccuracyEvaluator
from fsga.ml.models import ModelWrapper

# Load data
X_train, X_test, y_train, y_test, _ = load_dataset('iris', split=True)

# Setup
model = ModelWrapper('rf', n_estimators=50, random_state=42)
evaluator = AccuracyEvaluator(X_train, y_train, X_test, y_test, model)

# Run GA
from fsga.selectors.tournament_selector import TournamentSelector
from fsga.operators.uniform_crossover import UniformCrossover
from fsga.mutations.bitflip_mutation import BitFlipMutation

ga = GeneticAlgorithm(
    num_features=X_train.shape[1],
    evaluator=evaluator,
    selector=TournamentSelector(evaluator, tournament_size=3),
    crossover_operator=UniformCrossover(),
    mutation_operator=BitFlipMutation(probability=0.01),
    population_size=50,
    num_generations=100
)

results = ga.evolve()
print(f"Accuracy: {results['best_fitness']:.2%}")
print(f"Features: {results['best_chromosome'].sum()}/{X_train.shape[1]}")

Results

Dataset GA Accuracy Features Used Improvement
Breast Cancer 98.3% 14.3/30 (48%) +2.6% vs all features
Iris 97.7% 2.6/4 (65%) +2.7% vs all features
Wine 100% 7.2/13 (55%) +1.7% vs all features

Installation

pip install fsga

Or install from source:

git clone https://github.com/straightchlorine/fsga.git
cd fsga
pip install -e .

Documentation Structure

Project Status

Delivered: - GA framework with swappable operators - Experiment runner with baseline comparisons - Visualization plots - Statistical tests - Documentation

Future Extensions: - Parallel fitness evaluation - NSGA-II multi-objective optimization - Regression support - Deep learning integration

See Project Plan for roadmap.

License

MIT License (see pyproject.toml). A LICENSE file has not been added to the repository yet.