Glossary

Concise definitions of the terms used across the ml4 documentation.

In short

• Terms are grouped by area: core objects, feature engineering, the model roster, the post-fit layer and deployment.
• The two feature-engineering loci (GLOBAL and FOLD-LOCAL) are what make ml4's feature engineering leak-safe.
• The post-fit layer (calibration, threshold, intervals, baseline gate) is fitted after AutoGluon trains. Calibration, the threshold and the baseline gate can be re-decided without retraining.

Core

AutoGluon — the tabular AutoML engine ml4 delegates all model training, hyperparameter search, and ensembling to. ml4 wraps it with a feature-engineering recipe, a custom model roster, and a post-fit statistical layer.

AutoSupervised — the one-object supervised entry point. fit(train_data) takes a single DataFrame and returns a FittedPipeline.

Ml4Config — the supervised training configuration: a composite of per-area section dataclasses (data, engine, roster, constraints, …).

FittedPipeline — the deployable artifact: the trained predictor plus the frozen recipe and post-fit layers, applied at predict.

Preset — a named Ml4Config factory (quick_explore, high_accuracy, regulated_glassbox, …) that pre-sets a strategy. ml4 names no AutoGluon preset in the fit: the fit shape sets every AutoGluon lever itself.

Training policy — engine.training_policy: cross_validation (a bagged, cross-validated ensemble), train_validation (one deployed model validated on a reserved slice) or chronological_train_validation (the same, on a strictly-later slice of a temporal split). A temporal split needs the chronological policy, the chronological policy needs a temporal split, and a regulated run cannot use cross_validation.

Fit shape — how a run fits (whether it bags, over how many folds, whether it stacks, which rows the post-fit layers use), derived from the training policy, the split, the roster and the fold and stack budget by parleml_ml4.config.resolve_fit_shape; never stored in the config.

Feature engineering

Recipe — the frozen FeatureRecipe the planner produces from config + data; the exact, ordered feature-engineering program that is executed at fit and score. Stamped with its format version (ml4-recipe/2); a recipe from another version is refused.

GLOBAL locus — feature steps that are fit once, before folding, and never read the target (row-local or unsupervised-fitted). Run as one AutoGluon feature generator.

FOLD-LOCAL locus — target-aware steps (target encoders, CV-stacked text, group aggregates) re-fit once per bag fold on that fold's training rows, so a held-out row never influences its own encoding. This is where leakage safety is structural rather than scanned.

Leakage taxonomy — the classification of each feature step by how it touches the target (row_local, unsupervised_fitted, selection_instrumental, target_aware_stat, sample_altering), which pins it to a legal locus.

Target encoding — replacing a categorical with a target-derived statistic; ml4 ships seven standard category_encoders schemes (TE, WE, ME, JS, LO, CN, GM).

Model roster

Model spec — one entry of roster.models, and the unit everything roster-shaped is keyed by: a search space, an Optuna study, a trained model. A family with more than one shipped variant is variant-qualified ("GBM:dart"); a family with exactly one is named bare ("RF"). roster.STOCK_MODELS is the closed list.

Family — the library behind a spec (GBM = LightGBM, CAT = CatBoost, …), recovered from a spec by roster.family_of. The right key for anything that is a property of the library — the GPU build, the AutoGluon wrapper's weight-blindness — and the wrong key for anything that is a property of the tree shape.

Variant — the structural mode a booster trains in (gbdt / dart / rf; symmetric / depthwise / lossguide; gbtree / dart). Not a seed replica: AutoGluon's bagging already supplies those. Variants earn their roster slot by erring differently, which is the only thing a greedy weighted ensemble can exploit.

Prep options — GBM:extra and GBM:arith, the two halves of the LightGBM AutoGluon's extreme_quality preset calls Prep: its hyperparameters (extra-trees mode, small trees), and its arithmetic feature generator, which adds sums, differences, products and ratios of the run's columns to that one model's input. Opt-in, never in a preset; GBM:arith is experimental and cannot carry a monotonic constraint. See AutoGluon's Prep options.

Fold-local booster — a custom AutoGluon model (GBM / XGB / CAT) that runs the fold-local FE chain and binds native monotone constraints.

Glass-box model — an interpretable model family: GLM, scorecard, RuleFit, or EBM (GA2M). All four are regulated kinds.

Foundation challenger — a zero-shot in-context tabular model (TabICL v2) trained beside the stock roster and shown on the leaderboard as a challenger. It is never the default strongest model.

Regulated run — roster.regulated=True: a single, non-bagged, no-ensemble, deterministic transparent model, with the engine levers forced accordingly. It needs a train/validation training policy; cross_validation is refused.

Post-fit layer

Calibration — a probability re-map fit on out-of-sample scores and shipped only if it beats the raw probabilities. auto resolves to the one-vs-rest family (isotonic, or Platt below the rare-class floor) on a binary target and to temperature scaling on a multiclass one.

Threshold — the decision operating point tuned to the eval metric (F1 / F-β / accuracy-optimal); 0.5 is never chosen except as a degenerate fallback.

CQR — Conformalized Quantile Regression: the default regression prediction-interval method (with a split-conformal fallback).

Target transform — a regression target warp (log1p / sqrt / Yeo-Johnson) picked either by an original-units CV competition or, on a linear-only roster and on frames the competition cannot run on, by a model-free skewness gate. Every output is inverted back; the inverse is the plain analytic one unless Duan's smearing correction beat it out of sample, and measures.json records the route, the measured bias and which inverse serves (details).

Baseline gate — an honesty check comparing the model against the strongest feature-blind predictor (a constant or a chance-level draw); flags no_signal, worse_than_constant, or suspected_leakage.

Monotonic constraint — a pinned direction (+1 / -1) for a feature, either declared manually or auto-derived from the signed partial Spearman correlation on train (other numeric features held fixed).

Deployment

Bundle — the versioned directory FittedPipeline.save() writes / load() reads.

Treelite / GTIL — a compiler-free fast-inference accelerator for a single-LightGBM/XGBoost strongest model (the real_time preset), parity-guarded.

Airgapped export — a pure-Python scorer (delivery.export) that reproduces the whole pipeline with no AutoGluon at serve time.

Signal check — a standalone zero-shot probe that estimates whether a dataset has learnable signal before a full experiment.

Next: API reference · The training workflow