Skip to content

mltreelib

mltreelib logo

Tree-based machine learning with a Rust core. A data & feature layer that treats missing data as first-class, an adjustable benchmark generator with known ground truth, and a tree learner that handles categoricals and missing values natively — with rpart-faithful printing, SVG diagrams, and LightGBM-style monotone & interaction constraints.

  • Rust core (mltreelib-core), thin Python shell (PyO3)
  • Scikit-learn-shaped API — without the scikit-learn dependency
  • NA and NaN handled end-to-end: missing in → missing out; imputation opt-in
  • 51 TabArena-v0.1 benchmark datasets on board, plus a generator whose knobs control how much signal lives in group and hierarchy structure

ModelTree Piecewise Linear vs CART Step Function

flowchart TD
    subgraph RustCore["mltreelib-core (Rust Engine)"]
        direction LR
        Frame["Typed Frame & NA Policy"] --> SplitSearch["Greedy Split Search\n(SSE, Gini, Chow Gate, Oblique)"]
        SplitSearch --> LeafFit["Decoupled Leaf Solvers\n(Ridge OLS, Huber, Quantile, GLM)"]
    end
    subgraph PythonShell["mltreelib (Python FFI)"]
        direction LR
        ModelTree["ModelTree API"] & LeafControl["LeafControl"] & Metrics["Metrics & Diagnostic Curves"]
    end
    PythonShell <==>|"PyO3 Zero-Copy / Fast Memory"| RustCore

    classDef rustBox fill:#1e3a5f,stroke:#2b4c7e,stroke-width:2px,color:#ffffff;
    classDef pyBox fill:#0f5132,stroke:#198754,stroke-width:2px,color:#ffffff;
    classDef frameBox fill:#2c3e50,stroke:#34495e,stroke-width:2px,color:#ffffff;

    class Frame,SplitSearch,LeafFit rustBox;
    class ModelTree,LeafControl,Metrics pyBox;
uv sync                                          # venv + Rust extension (maturin)
uv run python examples/quickstart.py             # fit, print, visualize (34.3% RMSE reduction on Concrete)
uv run python examples/real_tree_studies.py wine  # after downloading Wine Quality
I want to… go to
install it Installation
see it work in 30 seconds Quickstart
choose a modeling task Practical modeling pathways
solve a real problem Real datasets and worked studies
change one part of a model Example library
wrangle data & encodings Examples — Data Processing
know every knob Examples — Model building & predictions
read the numbers Examples — Inspection
draw the tree Visualization
see how it compares Benchmarks — ELO
read technical deep dives & blogs Blog & Technical Insights
look up every parameter API reference

Explore our in-depth architectural and commercial case studies:

  1. Decision Trees for Credit Scoring Regulatory Compliance & Fintech: How tree paths, constraints and calibration support model review.
  2. Zero-Imputation Predictive Maintenance for Industrial Sensor Feeds Industrie 4.0 & Sensor Telemetry: A synthetic example of missing sensor values and tree routing.
  3. Uplift Modeling with Causal Trees Commercial Decision Optimization: How an honest causal tree compares targeting policies on a randomized held-out trial.
  4. Cubist Models and C5.0 Rules Core Systems ML: Linear leaves, extracted rules and the contracts still to check.
  5. Beyond Mean Squared Error: Distributional Decision Trees for Heavy Tails, Insurance, and Censored Data Actuarial & Risk Modeling: Tweedie compound Poisson-Gamma claims, Tobit censored log-likelihood, and heteroscedastic volatility regimes.
  6. Breaking the Step Function: Model Trees and Step Functions for Continuous Regression Systems ML & Physics-Informed ML: How linear leaves represent local trends and why fair comparisons need held-out data.
  7. Segregate First, Model Later: Two-Stage Hierarchical Survival Trees for Longitudinal Cohorts Actuarial & Retention Analytics: How decoupling ID-level cohort discovery from leaf-level Kaplan-Meier curves eliminates right-censoring bias and non-proportional hazard violations.

Try it right now

uv run python examples/quickstart.py fits a tree, prints it in rpart format and writes an SVG diagram. The end-to-end walkthrough then does the same on 10,000 real bank customers.

The research program, roadmap and contribution rules live in the repository (README.md, AGENTS.md).