Skip to content

Release notes

In short

  • mli 0.8.0 is the first release. It ships inside ParleML 0.8.0 and powers each experiment's Interpretation tab.
  • The installed version is parleml_mli.__version__.
  • mli explains fitted models; it does not train them. Models usually come from ml4.

0.8.0 — ships with ParleML 0.8.0

Profiling

  • fit_profiler profiles a fitted model and returns one ProfilingResult. Optional analyses run only when their ProfilingConfig section is present. See Profiling.
  • Archetypes: HDBSCAN clusters of the rows you choose (target_mask), in tree-leaf or SHAP space, with signatures, similar cases and opt-in bootstrap stability.
  • Binary classification, multiclass (one-vs-rest per class) and regression.
  • Results save to one joblib and reload without refitting. Saved artifacts carry a schema version; an incompatible one is refused with StaleArtifactError. See Consuming saved results.

Supported models

  • Native adapters for CatBoost, LightGBM, XGBoost, scikit-learn forests, extra trees, gradient boosting, HistGradientBoosting and single trees, EBM, RuleFit and scikit-learn linear models and GLMs.
  • monitor_predictions covers any other scorer, or stored predictions alone. See Bring-your-own-model and Monitoring predictions.
  • ml4 models: the best readable base tree is explained, engineered features roll back to their source columns, and the deployed pipeline scores. See Working with ML4.

Rules

  • IF-THEN label rules mined through the model's trees: the positive miner, the sequential miner and an optional all-class rule list.
  • Every rule has Train numbers and, given held-out rows, Test numbers with a holds-up verdict.
  • Native categorical splits read as the original labels; missing values and unseen categories follow the model's own routing. See Rule extraction.

Global explanations

  • explain: partial dependence with an ICE band, ALE, SHAP interactions, a prediction surrogate, a residual surrogate, permutation importance, class separation (PSI / Information Value), compacity, fairness (AIR, ME, SMD) and an error tree.
  • Effect curves draw the model's top drivers and record why any column was skipped. See Global explanations.

Data and evaluation

  • compute_eda: distributions, associations, missing values, outliers, duplicates, alerts and a leakage scan with six finding kinds. No model needed. See EDA & leakage.
  • Held-out evaluation for binary, multiclass and regression models, with calibration, business value and a multiclass cost matrix.
  • compute_label_noise: a confident-learning review queue of likely mislabelled rows.
  • compute_trust_summary: nine diagnostic tiles; missing evidence shows yellow.
  • build_model_card: a structured model card for the app to render. See Evaluation, trust & model cards.

Per-case decision support

  • justify explains one row: the prediction, its SHAP drivers, its archetype, the rules it fires and its nearest neighbours. ICE, LOCO and an archetype counterfactual are optional.
  • justify_batch produces reason codes for a whole frame. See Decision support.

Other experiment kinds and charts

  • Views for ml4 forecasting, clustering, anomaly detection and association rules.
  • More than 100 interactive Plotly figure functions across all areas. See Interactive visualizations.

AI-written text

  • Archetype names, summaries, rule explanations and column descriptions are annotations, written only through an Ask the host supplies. mli never calls a model itself. See AI-written narration.

Next