Examples¶
In short
- Each example is a complete Python walkthrough, from a fitted model to its explanations.
- Start with Explain an ml4 model if your model comes from ml4, or Bring-your-own-model if it comes from elsewhere.
- Every example uses one of two calls:
fit_profilerfor a supported model, ormonitor_predictionsfor any other scorer. - The examples assume the named input datasets exist; swap in your own table.
ml4 builds models; mli explains them. Most examples train a plain library model to keep the focus on mli. A model trained with ml4 works the same way once you have picked the tree to explain.
Walkthroughs¶
| Example | What it covers |
|---|---|
| Explain an ml4 model | Profile the best readable tree of an ml4 pipeline, in business column names, and explain a new row with the deployed pipeline |
| Binary classification (fraud) | Archetypes, rules, explanations, EDA, held-out evaluation, trust and a model card on a fraud classifier |
| Multiclass classification | One-vs-rest archetypes and rules, class-specific results, trust and a model card |
| Regression cohorts | Profile a high-value tail through target_mask, with per-archetype residual diagnostics |
| Bring-your-own-model | Native and lite paths, depending on whether mli has an adapter for your model |
The two entry points¶
| Call | When | What you get |
|---|---|---|
fit_profiler(model=..., X=..., y=..., config=...) |
Your model has a native adapter: CatBoost, LightGBM, XGBoost, scikit-learn ensembles and linear models, HistGradientBoosting, EBM, RuleFit, including the trees inside ml4 | Native profiling and the optional stages you request. SHAP, geometry and rule support vary by family |
monitor_predictions(...) |
Any other scorer, or predictions only | The lite path: surrogate, error tree, archetype-lite, metrics, drift and fairness |
A supervised run builds one ProfilingConfig and calls
fit_profiler(..., config=...) once. Report files go to ReportConfig(output_dir=...). Save a
result with result.save(...) and reload it with ProfilingResult.load(...); to read a saved
result without refitting, see Consuming saved results.
ml4 clustering, anomaly and forecasting results are drawn through parleml_mli.unsupervised and
parleml_mli.forecasting instead of fit_profiler. See
Interactive visualizations.