dxlearn 1.0.0

Genetic Algorithm Driven AutoML Framework for sklearn-compatible classification pipelines

dxlearn — Genetic Algorithm Driven AutoML Framework

dxlearn is a production-grade, research-ready AutoML Python package that discovers optimal classification pipelines using a grammar-constrained Genetic Algorithm (GA) built strictly on top of scikit-learn.

Features

• Grammar-constrained search: Pipelines follow <OptionalPreprocessor> <Scaler> <Classifier>.
• Multi-objective fitness: Accuracy, fit time, predict time, and complexity (scalarized for selection).
• sklearn-compatible API: fit, predict, predict_proba, score, get_params, set_params.
• Deterministic & reproducible: Optional seeded RNG and fitness caching.
• Extensible: Base abstractions for regression, NSGA-II, and distributed GA (future).

Requirements

• Python 3.11+
• numpy, scikit-learn, joblib

Installation

pip install -e .
# With dashboard (FastAPI + uvicorn):
pip install -e ".[dashboard]"

Quick Start

from dxlearn import DXClassifier
from sklearn.datasets import load_iris
from sklearn.model_selection import train_test_split

X, y = load_iris(return_X_y=True)
X_train, X_test, y_train, y_test = train_test_split(X, y, random_state=42)

model = DXClassifier(
    population_size=30,
    generations=20,
    cv=5,
    alpha=1.0,
    beta=0.2,
    gamma=0.01,
    max_runtime=600,
    verbose=2,
    n_jobs=-1,
    deterministic=True,
    random_state=42,
)

model.fit(X_train, y_train)
preds = model.predict(X_test)
print(model.score(X_test, y_test))
print(model.best_pipeline_)
print(model.best_score_)

# Optional: launch analytical dashboard
model.dashboard()  # requires pip install dxlearn[dashboard]

Pipeline Grammar (v1)

• OptionalPreprocessor: None | PCA | SelectKBest | PolynomialFeatures | VarianceThreshold
• Scaler: StandardScaler | MinMaxScaler | RobustScaler
• Classifier: LogisticRegression | RandomForestClassifier | GradientBoostingClassifier | SVC | KNeighborsClassifier | DecisionTreeClassifier

Fitness

Multi-objective vector: (accuracy, fit_time, predict_time, complexity).
Scalarized for selection: α·accuracy − β·fit_time − γ·complexity − δ·predict_time (default weights: α=1, β=0.2, γ=0.01).

Dashboard

With pip install dxlearn[dashboard], calling model.dashboard() starts a FastAPI server at http://127.0.0.1:8000 with:

• Generation evolution curves (best fitness, best accuracy)
• Accuracy vs time scatter
• Mean fitness over generations
• Best metrics summary

Package Layout

dxlearn/
├── base/           # BaseSearch, EvolutionarySearch, BaseDXEstimator
├── encoding/       # Grammar, tree, node (pipeline representation)
├── operators/      # Selection, crossover, mutation
├── search_space/   # Registry (scalers, preprocessors, classifiers)
├── evaluation/     # Evaluator, Objectives, Scalarizer
├── engine/         # GeneticSearch
├── dashboard/      # FastAPI dashboard (optional)
├── dxclassifier.py # Public API
└── config.py       # Defaults

License

MIT.