Practical modeling pathways¶
Start with a decision. Each pathway uses a real cited dataset, shows the complete fitting code, reads empirical metrics against baseline benchmarks, and provides transparent decision inspection.
flowchart TD
Start["What is your core machine learning task?"] --> Target{"What is the target response type?"}
Target -- "Continuous Numeric" --> Dist{"Is the target heavy-tailed, censored, or non-normal?"}
Dist -- "Standard Gaussian / Linear" --> G_Reg["Regression Pathway\n(Piecewise Linear ModelTree)"]
Dist -- "Heavy-tailed / Zero-inflated / Censored" --> G_Dist["Distributional Leaves Suite\n(Gamma, Poisson, Tweedie, Tobit)"]
Target -- "Binary / Multi-Class Categorical" --> G_Cla["Classification Pathway\n(Glass-Box Credit Distress & Review Queue)"]
Target -- "Time-to-Event with Right Censoring" --> G_Surv["Survival Pathway\n(Log-Rank Splits & Nelson-Aalen Corridors)"]
Target -- "Randomized Treatment / A/B Intervention" --> G_Uplift["Uplift Pathway\n(Honest Causal Trees & CATE Targeting)"]
Target -- "Unlabeled Observations (Unsupervised)" --> Unsup{"Are you isolating outliers or finding segments?"}
Unsup -- "Discover Homogeneous Cohorts" --> G_Clust["Clustering Pathway\n(Predictive Clustering Trees with Rules)"]
Unsup -- "Surface Anomalous Telemetry" --> G_Anom["Anomaly Detection Pathway\n(Extended Isolation Forest without Ghosts)"]
classDef startBox fill:#1e293b,stroke:#475569,stroke-width:2px,color:#ffffff;
classDef gateBox fill:#b45309,stroke:#d97706,stroke-width:2px,color:#ffffff;
classDef regBox fill:#1e3a5f,stroke:#2b4c7e,stroke-width:2px,color:#ffffff;
classDef distBox fill:#4c1d95,stroke:#6d28d9,stroke-width:2px,color:#ffffff;
classDef claBox fill:#9a3412,stroke:#c2410c,stroke-width:2px,color:#ffffff;
classDef survBox fill:#065f46,stroke:#059669,stroke-width:2px,color:#ffffff;
classDef upliftBox fill:#0f766e,stroke:#14b8a6,stroke-width:2px,color:#ffffff;
classDef clustBox fill:#0369a1,stroke:#0284c7,stroke-width:2px,color:#ffffff;
classDef anomBox fill:#881337,stroke:#9f1239,stroke-width:2px,color:#ffffff;
class Start startBox;
class Target,Dist,Unsup gateBox;
class G_Reg regBox;
class G_Dist distBox;
class G_Cla claBox;
class G_Surv survBox;
class G_Uplift upliftBox;
class G_Clust clustBox;
class G_Anom anomBox;
| Decision | Path | Dataset | What to check |
|---|---|---|---|
| Estimate a price | Regression | OpenML diamonds | Test errors and residuals |
| Predict financial distress | Classification | Give Me Some Credit | Ranking, probability and threshold cost |
| Group wine by chemistry | Clustering | UCI Wine Quality via OpenML | Leaf profiles and refit stability |
| Prioritize unusual readings | Anomaly detection | Scania APS Failure / Shuttle | Precision at a fixed review budget |
| Estimate survival over time | Survival | FLCHAIN | Censoring-aware error by horizon |
| Decide who gets an email | Uplift | Hillstrom experiment | Policy value against treat-all and treat-none |
How to use the documentation¶
- Run one pathway. Get its dataset and follow the analysis from split through evaluation.
- Change one part. The example library covers missing routing, alternative leaves, calibration diagnostics, retraining and other focused variations.
- Review the decision. Use the evaluation guide, visualization catalog and serving contract.
The dataset setup page lists the downloads. Synthetic checks sit after the practical workflows. They make known effects and missing-data behavior easier to test.