Glossary¶
Terms used throughout the ParleML MLI documentation.
Profiling — the core MLI run (fit_profiler) that explains a fitted model:
archetypes plus requested optional rules, explanations and EDA; evaluation needs
labelled splits, and trust is derived separately from the populated result.
ProfilingResult — the hub object every artifact hangs off (result.rules_,
result.explanations_, result.eda_, …). Save/load as a single joblib.
Archetype — a data-driven cluster of rows the model treats similarly, built
in attribution (SHAP / leaf) space. Summarized in result.summary_df.
target_mask — the boolean slice of rows the archetype profiling describes
(e.g. the positives, or a high-value tail).
SHAP — per-feature contributions to a prediction; MLI uses exact tree SHAP on native models.
Leaf-space vs SHAP-space — archetypes built from tree-leaf co-occurrence (leaf-space) vs from SHAP vectors (SHAP-space). Rule extraction is a separate capability: some SHAP-space families also have tree-rule extractors.
PDP / ICE — Partial Dependence (average effect of a feature) / Individual Conditional Expectation (per-row curves).
ALE — Accumulated Local Effects, a correlation-robust alternative to PDP.
Permutation importance — importance measured by the metric drop when a feature is shuffled.
PSI — Population Stability Index between two samples. Reference-versus-current PSI measures monitoring drift; target-versus-reference PSI describes class separation.
Fairness (AIR / ME / SMD) — Adverse Impact Ratio, Marginal Effect, and Standardized Mean Difference across groups.
Rules engine — mines human-readable IF-THEN label rules: patterns in the labels, found through the model's trees (see the Rules guide).
Lift / coverage / confidence — how much a rule concentrates the positive class / how many rows it fires on / P(positive | fired).
Train / Test numbers (rules) — a rule's numbers on the rows it was mined on (optimistic: the search picked the rule there) vs. the same numbers on held-out rows the rule search never saw — for an imported model, a held-back half of the reference sample (see Train and Test numbers).
Holds up — a rule's verdict on its Test numbers: with at least
min_test_rows (default 20) covered held-out rows, the Wilson lower bound of its
test precision exceeds the test base rate of its class; otherwise it does not
hold up, or has too few test rows below that floor.
Holdout basis — where a rule set's Test rows come from (HoldoutBasis.kind):
the fit's reserved validation slice plus the outer test, the outer test alone,
a seeded half of an imported model's reference sample (held out from rule
selection; the model's training rows are unknown), or none.
DSD (Decision Support Data) — the per-row explanation payload (drivers, matched archetype, neighbors, fired rules).
justify / CaseExplanation — produce the DSD for one row; optionally with
ICE / LOCO / counterfactual sensitivity.
LOCO — Leave-One-Covariate-Out, a per-row sensitivity measure.
Archetype counterfactual — a search along a path toward another archetype's medoid, reporting an archetype transition when found. It does not guarantee a globally minimal, feasible change or a flipped class prediction.
Trust summary — a traffic-light hero of governance tiles
(compute_trust_summary).
Model card / AutoDoc — a structured, shareable model document
(build_model_card).
EncodingRegistry / Ml4OriginRegistry — maps engineered features back to their
original columns so explanations read in business terms (backed by parleml_ml4.fe.naming).
monitor_predictions — the lite path for models without a native adapter
(surrogate / error tree / metrics / drift / fairness). See
Bring-your-own-model.
Narration — AI-written business-language text a host attaches as an annotation (see AI-written narration); the analysis itself is always deterministic.