Cost-weighted service loss regressor

This section documents the cost-weighted service loss (CWSL) regressor used in model selection workflows.

The CWSL regressor supports learning and ranking models based on asymmetric service risk preferences.

eb_evaluation.model_selection.cwsl_regressor

scikit-learn-style estimator wrapper for cost-aware model selection.

This module defines CWSLRegressor, a lightweight wrapper around ElectricBarometer that exposes a familiar fit / predict / score API.

The estimator selects among a set of candidate models using Cost-Weighted Service Loss (CWSL) as the selection criterion. Candidate models are trained using their native objectives (e.g., squared error), but selection is performed using asymmetric operational costs.

The primary cost preference can be summarized by the cost ratio:

\[ R = \frac{c_u}{c_o} \]

where \(c_u\) is the underbuild (shortfall) cost per unit and \(c_o\) is the overbuild (excess) cost per unit.

CWSLRegressor

scikit-learn-style estimator that selects among candidate models using CWSL.

This class wraps ElectricBarometer and exposes:

• fit(X, y): perform cost-aware model selection
• predict(X): predict using the selected best estimator
• score(X, y): return a sklearn-style score based on negative CWSL

Parameters:

Name	Type	Description	Default
models	dict[str, Any]	Mapping of candidate model name to an unfitted estimator implementing: fit(X, y) predict(X) Models may be scikit-learn regressors, pipelines, or EB adapters.	required
cu	float	Underbuild (shortfall) cost per unit. Must be strictly positive.	2.0
co	float	Overbuild (excess) cost per unit. Must be strictly positive.	1.0
tau	float	Tolerance parameter forwarded to ElectricBarometer for optional diagnostics (e.g., HR@tau).	2.0
training_mode	selection_only	Training behavior; currently ElectricBarometer supports selection-only mode.	"selection_only"
refit_on_full	bool	Refit behavior after selection. In holdout mode: if True, refit the winning model on train + validation. In CV mode: the winning model is refit on the full dataset.	True
selection_mode	(holdout, cv)	How selection is performed: "holdout": split internally using validation_fraction. "cv": perform K-fold selection inside ElectricBarometer.	"holdout"
cv	int	Number of folds when selection_mode="cv".	3
validation_fraction	float	Fraction of samples used for validation when selection_mode="holdout". Must be in (0, 1).	0.2
random_state	int | None	Seed used for internal shuffling and CV splitting.	None

Attributes:

Name	Type	Description
selector_	ElectricBarometer | None	Underlying selector instance used in the most recent fit.
best_name_	str | None	Name of the winning model.
best_estimator_	Any | None	Fitted winning estimator.
results_	Any	Comparison table produced by ElectricBarometer (typically a pandas DataFrame).
validation_cwsl_	float | None	CWSL score of the winning model on validation (holdout) or mean CV.
validation_rmse_	float | None	RMSE score of the winning model on validation or mean CV.
validation_wmape_	float | None	wMAPE score of the winning model on validation or mean CV.
n_features_in_	int | None	Number of features observed during fit.

Notes

score returns negative CWSL to align with sklearn conventions (higher is better).

r_ property

Cost ratio.

Returns:

Type	Description
float	The ratio \(R = c_u / c_o\).

fit(X, y, sample_weight=None)

Fit CWSLRegressor on (X, y) by delegating selection to ElectricBarometer.

Parameters:

Name	Type	Description	Default
X	array-like of shape (n_samples, n_features)	Feature matrix.	required
y	array-like of shape (n_samples,)	Target vector.	required
sample_weight	numpy.ndarray of shape (n_samples,)	Optional per-sample weights. In CV mode these are passed through to ElectricBarometer so validation folds are cost-weighted. In holdout mode, they are passed through to ElectricBarometer (behavior depends on EB).	None

Returns:

Type	Description
CWSLRegressor	Fitted instance.

Raises:

Type	Description
ValueError	If X/y shapes are incompatible or sample_weight length mismatches.

predict(X)

Generate predictions from the selected best estimator.

Parameters:

Name	Type	Description	Default
X	array-like of shape (n_samples, n_features)		required

Returns:

Type	Description
numpy.ndarray of shape (n_samples,)	Predicted values.

Raises:

Type	Description
RuntimeError	If called before fit.

score(X, y, sample_weight=None)

Compute a sklearn-style score using negative CWSL.

Parameters:

Name	Type	Description	Default
X	array-like of shape (n_samples, n_features)	Feature matrix.	required
y	array-like of shape (n_samples,)	Target vector.	required
sample_weight	numpy.ndarray of shape (n_samples,)	Optional per-sample weights passed to CWSL.	None

Returns:

Type	Description
float	Negative CWSL on the provided data (higher is better).

get_params(deep=True)

Minimal sklearn-compatible get_params implementation.

Parameters:

Name	Type	Description	Default
deep	bool	Included for sklearn compatibility. Nested parameter extraction is not performed in this lightweight implementation.	True

Returns:

Type	Description
dict[str, Any]	Parameter mapping.

set_params(**params)

Minimal sklearn-compatible set_params implementation.

Parameters:

Name	Type	Description	Default
**params		Parameters to set on this instance.	{}

Returns:

Type	Description
CWSLRegressor	Updated instance.

Raises:

Type	Description
ValueError	If an invalid parameter name is provided.