Example Library¶
The example library is organized into 9 authoritative suites. Each suite grounds its implementation in a real cited benchmark dataset, explains the theoretical rationale and domain use case, provides architectural Mermaid flowcharts, and generates high-resolution scikit-learn style diagnostic figures locked to the #00897b Teal / #f4511e Deep Orange palette.
The 9 Consolidated Example Suites¶
-
Suite 1: Core Tree Architecture & Growth Knobs
Two-Control decoupling (
ModelTree\(\leftrightarrow\)LeafControl), 0D leaf stats, and Chow structural break stopping on Yeh Concrete. -
Suite 2: Structural Constraints & Honest Inference
Dual monotone split and slope constraints, interaction isolation, and sample-split honesty (\(S^{\text{tr}}\) vs \(S^{\text{est}}\)) on Healthcare data.
-
Suite 3: Native Missing Data Routing
Learned
missing_leftdirection and surrogate splits (maxsurrogate=N) for zero-imputation telemetry triage on Scania APS. -
Suite 4: Distributional & Censored Leaf Families
Gamma, Poisson, Tweedie, and Tobit censored MLE leaves with 60%+ OLS attenuation correction across actuarial risk datasets.
-
Suite 5: Dynamic Leaf Retraining & Adaptation
Microsecond warm-start leaf retraining, Huber loss breakdown resistance (25.8 vs 54.3 MPa under sensor glitch), and quantile corridors.
-
Suite 6: Statistical, Metric & Oblique Splitting
Chow structural break gate, oblique ridge splits, Gain Ratio, Twoing, SDR, and Quinlan rule extraction on Housing and Wine.
-
Suite 7: Time Series Feature Engineering & Forecasting
Pure Rust calendar extraction, entity-grouped calendar day lags, additive decomposition, and out-of-time trend extrapolation.
-
Suite 8: Unsupervised Trees: Clustering & Anomaly Triage
Predictive Clustering Trees (PCT) with rule extraction and Extended Isolation Forest (EIF) with oblique cuts for multi-target grouping.
-
Suite 9: Survival & Competing Risks Trees
Gray's subdistribution test, weighted log-rank criteria, \(O(N)\) streaming martingale scans, and Aalen-Johansen CIF simplex invariant.
-
Leaf Solver Benchmarks: 7 Optimization Algorithms
Pareto frontier of fit latency vs. accuracy across Analytical (
irls), Quasi-Newton (lbfgs), Coordinate (ccd), and Evolutionary (shade,cmaes,ga,nes) solvers.
Detailed Mechanics Reference Table¶
| Suite | Focus Area | Benchmark Dataset | Key Mechanics Covered |
|---|---|---|---|
| Suite 1: Core Tree Architecture & Growth Knobs | Core Systems | Yeh Concrete (1,030 rows) | Two-Control decoupling (ModelTree <-> LeafControl), 0D leaf stats, Chow hypothesis stopping. |
| Suite 2: Structural Constraints & Honest Inference | Physics & Governance | Healthcare Expenses (1,338 rows) | Dual monotone split and slope constraints, interaction isolation, sample-split honesty (\(S^{\text{tr}}\) vs \(S^{\text{est}}\)). |
| Suite 3: Native Missing Data Routing | Industrial IoT | Scania APS Failure (76,000 rows) | Learned missing_left direction, surrogate splits (maxsurrogate=N), zero-imputation telemetry triage. |
| Suite 4: Distributional & Censored Leaf Families | Actuarial & Risk | Credit-G, Healthcare, Affairs, Mroz | Gamma, Poisson, Tweedie, and Tobit censored MLE leaves with 60%+ OLS attenuation correction. |
| Suite 5: Dynamic Leaf Retraining & Adaptation | Streaming & Systems | Yeh Concrete (1,030 rows) | Microsecond warm-start leaf retraining, Huber breakdown resistance (25.8 MPa vs 54.3 MPa under glitch), quantile corridors. |
| Suite 6: Statistical, Metric & Oblique Splitting | Split Optimization | Boston/Miami Housing, Wine Quality | Chow structural break gate, oblique ridge splits, Gain Ratio, Twoing, SDR, and Rule extraction. |
| Suite 7: Time Series Feature Engineering & Forecasting | Temporal Econometrics | Multi-store retail demand panel | Pure Rust calendar extraction, entity-grouped calendar day lags, additive decomposition, and out-of-time trend extrapolation. |
| Suite 8: Unsupervised Trees: Clustering & Anomaly Triage | Discovery & Triage | Wine Quality & 2D Mixture | Predictive Clustering Trees (PCT) with rule extraction, Extended Isolation Forest (EIF) with oblique cuts. |
| Suite 9: Survival & Competing Risks Trees | Clinical & Churn | Mayo PBC, Veterans, FLCHAIN | Gray's test, Tarone-Ware/Peto/Martingale, Aalen-Johansen CIF simplex invariant (\(\sum F_c + S = 1\)), Uno IPCW C. |
| Leaf Solver Benchmarks | Optimization Systems | Yeh Concrete (1,030 rows) | Analytical Cholesky, Quasi-Newton L-BFGS, Cyclic Coordinate Descent, and Evolutionary Solvers (SHADE, CMA-ES, GA, NES). |
Architectural Workflow Navigator¶
flowchart TD
Start["What modeling challenge do you need to solve?"] --> Q1{"What is the target / learning setting?"}
Q1 -- "Continuous Tabular Response" --> A1["Suite 1: Core Tree Architecture\n(Linear Leaves & Growth Knobs)"]
Q1 -- "Preserve Domain Laws / Fairness" --> A2["Suite 2: Structural Constraints\n(Monotone & Honest Inference)"]
Q1 -- "Intermittent Telemetry / Missing Inputs" --> A3["Suite 3: Native Missing Data\n(Learned Routes & Surrogates)"]
Q1 -- "Heavy Tails / Zero Inflation / Censoring" --> A4["Suite 4: Distributional Leaves\n(Tweedie, Gamma, Tobit)"]
Q1 -- "Sub-Millisecond Adaptation & Online Streams" --> A5["Suite 5: Dynamic Leaf Retraining\n(Warm-Start Hierarchy & Huber)"]
Q1 -- "Structural Parameter Breaks & Oblique Cuts" --> A6["Suite 6: Statistical & Metric Splitting\n(Chow Test & Oblique Scans)"]
Q1 -- "Autoregressive Demand & Seasonality" --> A7["Suite 7: Time Series Forecasting\n(Calendar Lags & Additive Decomp)"]
Q1 -- "Unsupervised Grouping / Outlier Triage" --> A8["Suite 8: Unsupervised Trees\n(PCT Clustering & EIF Anomaly)"]
Q1 -- "Right-Censored / Competing Multi-Cause" --> A9["Suite 9: Survival & Competing Risks\n(Gray's Test, Aalen-Johansen CIF)"]
classDef startBox fill:#1e293b,stroke:#475569,stroke-width:2px,color:#ffffff;
classDef gateBox fill:#b45309,stroke:#d97706,stroke-width:2px,color:#ffffff;
classDef suiteBox fill:#1e3a5f,stroke:#2b4c7e,stroke-width:2px,color:#ffffff;
classDef survBox fill:#065f46,stroke:#059669,stroke-width:2px,color:#ffffff;
class Start startBox;
class Q1 gateBox;
class A1,A2,A3,A4,A5,A6,A7,A8 suiteBox;
class A9 survBox;
Reference Guides & End-to-End Studies¶
- For complete industry case studies on regulatory credit scoring, sensor maintenance, and causal marketing, visit the Practical Modeling Pathways.
- For complete function signatures and parameter specifications, visit the API Reference.