daisugi

daisugi: not-so-well-known tree machines daisugi is an R package collecting experimental, obscure, and emerging tree-based machine learning methods. Rather than reproducing the mainstream boosting ecosystem, daisugi focuses on: probabilistic forests hybrid boosting systems online learners interpretable ensembles experimental tree architectures research-oriented methods rarely exposed to R users Note: popular libraries such as XGBoost, LightGBM, & CatBoost are included but modified to use daisugi defaults. Included Machines v0.0.10 Status Algorithm Focus ✅ Boulevard stochastic gradient boosting ✅ ChimeraBoost numba-accelerated boosting ✅ CycGBM distributional boosting ✅ FIGS interpretable greedy tree sums ✅ Conditional Trees unbiased recursive partitioning ✅ EBM interpretable additive boosting ✅ Evolutionary Trees genetic tree optimization ✅ Extreme eXtreme gradient boosted forest 🚧 FairGBM fairness-aware boosting ✅ GRANDE differentiable tree ensembles ✅ Langevin Catboost with SGLB ✅ Linear LightGBM with GBDT-PL ✅ MIBoost multiple imputation boosting ✅ MorphBoost adaptive boosting structures 🚧 MSBoost multi-stage boosting ✅ NGBoost natural gradient prediction ✅ NRGBoost energy-based generative boosting ✅ Perpetual continual tree learning ✅ SkopeRules interpretable rule lists ✅ SnapBoost heterogeneous boosting systems ✅ STreeD optimal sparse trees ✅ TAO tree alternating optimization ✅ WildWood randomized online forests ✅ Yggdrasil scalable tree ecosystems Dropped engines Algorithm Reason scikit-garden abandoned upstream since 2017, does not build on modern Python PkBoost no Linux/macOS wheels, sdist requires Rust toolchain KTBoost abandoned upstream since 2021; author moved to GPBoost (#23) Blocked engines FairGBM and pysortd break on scikit-learn >= 1.6 due to removed APIs (force_all_finite, _validate_data). These need vendoring with patches or upstream fixes before they can be added. Installation You can install the development version of daisugi from GitHub: # install.packages("pak") pak::pak("frankiethull/daisugi") Philosophy daisugi explores tree systems outside the conventional gradient boosting canon. Many included methods emphasize: uncertainty estimation heterogeneous base learners recursive partition hybrids online adaptation probabilistic outputs alternative split mechanics interpretable ensemble structures The package acts as both: a practical modeling toolkit a curated collection of unconventional tree algorithms Example library(daisugi) model <- grow_yggdrasil_trees( x = iris[, 1:4], y = iris$Species ) harvest_yggdrasil_trees(model, iris[, 1:4]) Further Documentation Get Started with a Classification Task Or maybe you prefer Regression Learn more about daisugi via the Glossary. A full list of models & their methods can be found in the Package Index daisugi also contains tidymodels implementations for Explainable & Yggdrasil Boosting Machines Notes on each verstion at the Changelog