`monitor_predictions`¶

The model-agnostic bring-your-own-predictions entry point. Given either a generic model (any predict_proba / decision_function / predict scorer) or a precomputed predictions array, it runs the lite pipeline — surrogate tree, high-error cohort tree, archetype-lite segments, holdout metrics, drift (PSI) and model-free fairness — treating the model as an opaque scorer. See the BYO-model example.

In short

monitor_predictions treats the model as an opaque scorer, or takes predictions= alone.
It returns a ProfilingResult whose backend is lite_model or lite_predictions.
For a model mli has an adapter for, call fit_profiler instead.
The positive class and the decision cut are resolved once and stored on result.target_ and result.threshold_.

Prefer fit_profiler for natively supported models

For a model mli has an adapter for — CatBoost / LightGBM / XGBoost / scikit-learn ensembles (leaf-space), or HistGradientBoosting / DecisionTree / EBM / RuleFit / scikit-learn linear & GLM families (SHAP-space) — call fit_profiler directly for native profiling and requested optional analysis. Geometry and rule support vary by family; a binary y profiles the model's positive class (classes_[1]) without a mask. See Supported models. Tasks: binary, multiclass, and regression (task_type=).

Positive class and decision threshold

Both entry points resolve a binary run's positive class once (positive_class=, else the model's classes_[1], else True / 1 / the last of two sorted labels — the last with a warning) and its decision cut once (served_threshold= for a deployed operating point, else decision_threshold=: a float, "f1" / "f<beta>" to tune, or None — the default — for the model's own decision). The answers are stored on result.target_ and result.threshold_. positive_class= names a label by ml4's rule — string or numeric form, so "True" names True and "1.0" names 1.0 — and a y showing a single unconventional label (say, all "no") with nothing naming the positive class is refused rather than guessed.

monitor_predictions ¶

monitor_predictions(
    *,
    X: DataFrame,
    y: ndarray | Series | None = None,
    model: Any | None = None,
    predictions: ndarray | Series | None = None,
    X_valid: DataFrame | None = None,
    y_valid: ndarray | Series | None = None,
    predictions_valid: ndarray | Series | None = None,
    X_test: DataFrame | None = None,
    y_test: ndarray | Series | None = None,
    predictions_test: ndarray | Series | None = None,
    reference_df: DataFrame | None = None,
    sensitive_features: DataFrame | Series | None = None,
    task_type: TaskType = BINARY,
    mode: Literal[
        "surrogate", "error", "archetype_lite", "all"
    ] = "all",
    decision_threshold: float | str | None = None,
    served_threshold: float | None = None,
    positive_class: Any = None,
    surrogate_engine: "EngineKey" = "lgbm",
    random_state: int = 42
) -> ProfilingResult

Interpret a model's predictions — the model-agnostic lite path.

The standalone monitoring / "bring-your-own-predictions" entry point: given either a generic model (any object exposing predict_proba / decision_function / predict) or a precomputed predictions array, build a surrogate decision tree, high-error cohort tree, archetype-lite segments, holdout metrics, drift (PSI) and model-free fairness — without ever needing the model's internals. This is the standalone / external-model-monitoring path for scorers mli has no native adapter for.

For models mli recognises natively (CatBoost / LightGBM / XGBoost / sklearn ensembles / trees / HGB) prefer :func:parleml_mli.fit_profiler: it delivers the full surface (exact SHAP, leaf-space archetypes, rules, PDP/ICE, counterfactuals). monitor_predictions treats the model as an opaque scorer and works for any estimator.

Parameters:

Name	Type	Description	Default
`X`	`DataFrame`	Feature matrix and (optional) ground truth. `y` unlocks the error-tree branch and holdout metrics; without it the error tree is skipped (`capabilities.error_tree` stays `False`).	required
`y`	`DataFrame`	Feature matrix and (optional) ground truth. `y` unlocks the error-tree branch and holdout metrics; without it the error tree is skipped (`capabilities.error_tree` stays `False`).	required
`model`	`Any \| None`	Any estimator exposing `predict_proba` / `decision_function` / `predict`. Mutually exclusive with `predictions`.	`None`
`predictions`	`ndarray \| Series \| None`	Score array aligned with `X` (positive-class probability for binary; raw target for regression). Mutually exclusive with `model`.	`None`
`X_valid`	`DataFrame \| None`	Holdout splits for the evaluation report. In the predictions-only branch supply the matching `predictions_` alongside `X_` / `y_*`; missing scores simply leave that split off the report.	`None`
`y_valid`	`DataFrame \| None`	Holdout splits for the evaluation report. In the predictions-only branch supply the matching `predictions_` alongside `X_` / `y_*`; missing scores simply leave that split off the report.	`None`
`predictions_valid`	`DataFrame \| None`	Holdout splits for the evaluation report. In the predictions-only branch supply the matching `predictions_` alongside `X_` / `y_*`; missing scores simply leave that split off the report.	`None`
`X_test`	`DataFrame \| None`	Holdout splits for the evaluation report. In the predictions-only branch supply the matching `predictions_` alongside `X_` / `y_*`; missing scores simply leave that split off the report.	`None`
`y_test`	`DataFrame \| None`	Holdout splits for the evaluation report. In the predictions-only branch supply the matching `predictions_` alongside `X_` / `y_*`; missing scores simply leave that split off the report.	`None`
`predictions_test`	`DataFrame \| None`	Holdout splits for the evaluation report. In the predictions-only branch supply the matching `predictions_` alongside `X_` / `y_*`; missing scores simply leave that split off the report.	`None`
`reference_df`	`DataFrame \| None`	Drift baseline (defaults to `X_valid` when supplied).	`None`
`sensitive_features`	`DataFrame \| Series \| None`	Group memberships aligned with `X` for model-free per-group rates (stored on `result.meta['fairness_lite']`). The adverse-impact ratio reads against the highest-rate group of at least :attr:`parleml_mli.FairnessConfig.min_group_size` rows, as the native path's :func:`~parleml_mli.explanations.compute_fairness` does.	`None`
`task_type`	`TaskType`	Lite-pipeline knobs (see :func:`parleml_mli.fit_profiler` for the shared semantics).	`BINARY`
`mode`	`TaskType`	Lite-pipeline knobs (see :func:`parleml_mli.fit_profiler` for the shared semantics).	`BINARY`
`surrogate_engine`	`TaskType`	Lite-pipeline knobs (see :func:`parleml_mli.fit_profiler` for the shared semantics).	`BINARY`
`random_state`	`TaskType`	Lite-pipeline knobs (see :func:`parleml_mli.fit_profiler` for the shared semantics).	`BINARY`
`decision_threshold`	`float \| str \| None`	Binary decision cut, resolved once by :func:`parleml_mli.core.threshold.resolve_threshold` on the (scored or supplied) predictions: `None` (default) keeps the model's own decision (`0.5` on the positive-class probability), a `float` is used as given, `"f1"` / `"f<beta>"` tunes a cut against `y`. Stored on `result.threshold_`.	`None`
`served_threshold`	`float \| None`	The deployed model's published cut; wins over `decision_threshold`.	`None`
`positive_class`	`Any`	The positive label of a binary `y`, resolved once by :func:`parleml_mli.core.target.resolve_target` (`model.classes_[1]` when the model has classes, else `True` / `1` / the last sorted label) and stored on `result.target_`. `predictions` must be this class's probability.	`None`

Returns:

Type	Description
`ProfilingResult`	Carries a :class:`~parleml_mli.result.Capabilities` stamp (`backend` is `"lite_model"` or `"lite_predictions"`) so consumers can grey out tabs the lite path cannot power.

monitor_predictions¶

monitor_predictions ¶

Next¶

`monitor_predictions`¶