glazzbocks 0.1.4

A glass-box machine learning toolbox for interpretable pipelines

Machine Learning Toolbox

This repository provides a collection of modular, production-ready tools for building, evaluating, and interpreting machine learning models. Each tool is built with a focus on clean design, extensibility, and practical utility.

---

Philosophy: Explainable, Transparent, Glass-Box ML

Unlike traditional AutoML tools that act as "black boxes", this toolbox is designed to be a glass box — every decision, transformation, and output is fully transparent and controllable.

This toolbox is ideal for:

• Data scientists who value interpretable models
• Regulated industries requiring auditability
• Educators and learners exploring machine learning foundations
• Teams that prioritize trust and understanding over automation

We emphasize explainability through SHAP, feature importances, coefficients, and detailed diagnostics at every stage.

---

Contents

Module	Description
DataExplorer.py	Exploratory Data Analysis (EDA) and VIF calculation
ML_pipeline.py	Full preprocessing + modeling pipeline with cross-validation & diagnostics
ModelInterpreter.py	SHAP, feature importance, and coefficient visualizations

---

Module Overviews

DataExplorer.py

A lightweight class for quick exploratory analysis:

• Displays dataset shape, dtypes, and missing values
• Plots target distribution (auto-detects regression vs classification)
• Correlation heatmap and Variance Inflation Factor (VIF)
• Returns median-imputed numeric-only DataFrame for diagnostics

ML_pipeline.py

A complete scikit-learn-based pipeline manager:

• Auto-detects numerical and categorical columns
• Builds preprocessing pipeline (scaling, imputation, encoding)
• Supports both regression and classification
• Cross-validation with metrics, ROC, F1-thresholds, and confusion matrix
• Built-in visualizations for:
  • Predicted vs. Actual
  • Residual plots
  • Error distribution
  • ROC curve and F1-threshold optimization

ModelInterpreter.py

Interpret model behavior post-training:

• Works with pipelines and standalone models
• Tree-based models: Feature importances
• Linear models: Coefficients (with optional plot)
• Universal SHAP summary plot (auto-handles pipelines)

---

Getting Started

Each module can be used independently. Example usage:

from ML_pipeline import MLPipeline
pipeline = MLPipeline()
X_train, X_test, y_train, y_test = pipeline.split_data(df, 'target')
pipeline.fit(X_train, y_train)
pipeline.plot_roc_curve(X_test, y_test)

Or for model interpretation:

from ModelInterpreter import ModelInterpreter
interpreter = ModelInterpreter(model, X_train, task='classification')
interpreter.shap_summary()

---

Requirements

• scikit-learn
• pandas, numpy
• matplotlib, seaborn
• shap
• statsmodels (for VIF)

---

Notes

• SHAP is optimized for tree-based models; linear models are also supported
• Pipelines handle preprocessing internally—no need to do it manually
• Modules follow sklearn conventions for compatibility and ease of use