Skip to content

Changelog

All notable changes to this project will be documented in this file.

Format: Keep a Changelog; versioning: SemVer.

[Unreleased]

[0.1.1] - 2026-07-08

Added

  • export_rules_html: self-contained interactive HTML rules explorer (feature dropdown, shape curves, rule table; no dependencies)
  • Documentation: detailed algorithm walkthrough with pipeline schema (user-guide/algorithm.md); the explorer embedded live on the visualization page and inline in the German Credit guidebook notebook

[0.1.0] - 2026-07-07

Added

  • Core estimators: sklearn-compatible FlagGAMClassifier and FlagGAMRegressor with univariate screening, Benjamini–Hochberg FDR control, automatic missing-value basis discovery, feature weighting, rule export, and per-observation attribution
  • Rule basis construction: threshold flags, categorical level flags, hinge and trend terms (regression), and missing-value indicators
  • Benchmark suite: repeated-split protocol with training-only tuning carves; method registry with runners for paper Tables 3 (classification AUROC), 4 (regression RMSE/R²), 5 (robustness to missing values and feature noise), 7 (FlagGAM ablations), and 8 (hyperparameter sensitivity); German Credit smoke-acceptance test
  • Dataset loaders with parquet caching: Pima, Breast Cancer, Heart, German Credit, Adult, Bank Marketing (classification); Ames, California Housing, Wine Quality (regression)
  • PD calibration extension: diagnostics (reliability curves, Brier score, ECE, calibration-in-the-large) and recalibration methods (Platt, isotonic, base-rate offset) with cross-fitting to prevent data leakage
  • Monotonicity constraints extension: sign-constrained additive heads for exact feature monotonicity in classification and regression, via box-constrained L-BFGS-B optimization
  • Fairness extension: group fairness metrics (demographic parity, equal opportunity, AUROC gap) and rule-level proxy audit with binarized-indicator association ranking
  • Visualization module: six matplotlib plots — feature shapes, basis importance, explanation waterfall, prediction reliability, fairness group metrics, protected-attribute association
  • Documentation: MkDocs site with API reference via mkdocstrings, getting-started guide, user guides for rules/screening, extensions, benchmarks, and visualization; German Credit credit-approval guidebook as interactive Jupyter notebook