glassbox-automl 1.0.0

Transparent, scratch-built AutoML library (NumPy-only core)

GlassBox-AutoML

A transparent, scratch-built Automated Machine Learning library with a NumPy-only math core.

Features

Module	Contents
EDA / Inspector	Mean, median, mode, std, skewness, kurtosis, Pearson matrix, IQR outliers, auto-typing
Preprocessing	SimpleImputer, MinMaxScaler, StandardScaler, OneHotEncoder, LabelEncoder
Models	LinearRegression, LogisticRegression, DecisionTree, RandomForest, GaussianNaiveBayes, KNearestNeighbors
Optimization	GridSearch, RandomSearch, KFoldCV
Evaluation	ClassificationMetrics (accuracy, precision, recall, F1, confusion matrix), RegressionMetrics (MAE, MSE, RMSE, R²)
AutoFit	End-to-end pipeline: EDA → Cleaning → Model Search → Explainability report

Quick Start

from glassbox import AutoFit
import numpy as np

# data = numpy array, last column = target
af = AutoFit(task="classification", target_col=-1, cv=5, time_budget=60)
report = af.fit(data, feature_names=["age", "income", "credit_score", "approved"])

print(af.explain())
predictions = af.predict(new_X)

Installation

pip install numpy
pip install -e .

Run the demo

python examples/autofit_demo.py

Run all tests

python tests/test_utils.py
python tests/test_preprocessing.py
python tests/test_models.py
python tests/test_optimization_eval.py

Architecture

glassbox/
├── autofit.py              # End-to-end AutoML orchestrator
├── eda/
│   └── inspector.py        # EDA: statistics, correlation, outliers, auto-typing
├── preprocessing/
│   ├── imputer.py          # SimpleImputer (mean/median/mode)
│   ├── scalers.py          # MinMaxScaler, StandardScaler
│   └── encoders.py         # OneHotEncoder, LabelEncoder
├── models/
│   ├── linear.py           # LinearRegression, LogisticRegression (gradient descent)
│   ├── tree.py             # DecisionTree (Gini / MSE)
│   ├── forest.py           # RandomForest (bagging + √features)
│   ├── naive_bayes.py      # GaussianNaiveBayes (Laplace smoothing)
│   └── knn.py              # KNearestNeighbors (Euclidean + Manhattan)
├── optimization/
│   ├── search.py           # GridSearch, RandomSearch
│   └── cross_validation.py # KFoldCV
└── evaluation/
    └── metrics.py          # ClassificationMetrics, RegressionMetrics

Design Principles

• Zero heavy dependencies — only NumPy for all math
• White-box — every model can explain its decisions
• WASM-ready — no C extensions, pure Python + NumPy
• Modular — every transformer implements fit(), transform(), fit_transform()

License

MIT