RegressionCV

class RegressionCV(BaseAutoCV, RegressorMixin, ExplainerMixIn):
    """
    Defines an auto regression tree, based on the bayesian optimization base class.
    """

    model_: XGBRegressor | CatBoostRegressor
    feature_importances_: NDArray[np.float64]
    explainer_: TreeExplainer

    @validate_call(config={"arbitrary_types_allowed": True})
    def __init__(
        self,
        metric: AcceptableRegression,
        cv: BaseCrossValidator,
        n_trials: NonNegativeInt,
        timeout: NonNegativeInt,
        config: RegressionCVConfig,
        backend: str = "xgboost",
    ) -> None:
        """
        Constructor for RegressionCV.

        Args:
            metric: Loss metric to use as base for the estimation process.
            cv: Splitter object to use when estimating the model.
            n_trials: Number of optimization trials to use when finding a model.
            timeout: Timeout in seconds to stop the optimization.
            config: Configuration to use when fitting the model.
            backend: Backend to use for the model. Either "xgboost" or "catboost".
        """
        super().__init__(metric, cv, n_trials, timeout)
        self.config = config
        self.backend = backend

    def explain(self, X: Inputs, **explainer_params) -> dict[str, NDArray[np.float64]]:
        """
        Explains the inputs.
        """

        check_is_fitted(self, "model_", msg="Model is not fitted.")

        if getattr(self, "explainer_", None) is None:
            self.explainer_ = TreeExplainer(self.model_, **explainer_params)

        return {
            "mean_value": self.explainer_.expected_value,
            "shap_values": self.explainer_.shap_values(self._validate_X(X)),
        }

    def fit(self, X: Inputs, y: GroundTruth, **fit_params) -> "RegressionCV":
        """
        Fits RegressionCV.

        Args:
            X: input data to use in fitting trees.
            y: actual targets for fitting.
        """
        self.feature_names_ = list(X.columns) if isinstance(X, pd.DataFrame) else []

        constraints = self.config.get_kwargs(self.feature_names_, backend=self.backend)

        if self.backend == "xgboost":
            estimator_type = partial(XGBRegressor, enable_categorical=True)
        elif self.backend == "catboost":
            estimator_type = partial(
                CatBoostRegressor, verbose=False, allow_writing_files=False
            )
        else:
            raise ValueError(
                f"Unknown backend: {self.backend}. Must be 'xgboost' or 'catboost'."
            )

        self.model_ = self.optimize(
            estimator_type=estimator_type,
            X=self._validate_X(X),
            y=self._validate_y(y),
            parameters=self.config.parameters,
            return_train_score=self.config.return_train_score,
            backend=self.backend,
            X_validation=fit_params.get("X_validation", pd.DataFrame()),
            y_validation=fit_params.get("y_validation", pd.Series()),
            **constraints,
        )
        self.feature_importances_ = self.model_.feature_importances_

        return self

    def predict(self, X: Inputs) -> Predictions:
        """
        Returns model predictions.
        """
        check_is_fitted(self, "model_", msg="Model is not fitted.")
        return self.model_.predict(self._validate_X(X))

    @property
    def scorer(self) -> tp.Callable[..., float]:
        """
        Returns correct scorer to use when scoring with RegressionCV.
        """
        metric_func = self._resolve_metric(regression_metrics)
        return make_scorer(metric_func, greater_is_better=False)

@dataclass(frozen=True, config={"arbitrary_types_allowed": True})
class RegressionCVConfig:
    """
    Available config to use when fitting a regression model.

    monotone_constraints: dictionary containing monotonicity direction allowed for each
        variable. 0 means no monotonicity, 1 means increasing and -1 means decreasing
        monotonicity.
    interactions: list of lists containing permitted relationships in data.
    n_jobs: Number of jobs to use when fitting the model.
    parameters: dictionary with distribution bounds for each hyperparameter to search
        on during optimization.
    return_train_score: whether to return the train score when fitting the model.
    quantile_alpha: Quantile alpha to use when fitting the model, if fitting a quantile
        model.
    """

    monotone_constraints: dict[str, int]
    interactions: list[list[str]]
    n_jobs: int
    parameters: OptimizerParams
    return_train_score: bool

    def get_kwargs(self, feature_names: list[str], backend: str = "xgboost") -> dict:
        """
        Returns parsed and validated constraint configuration for a RegressionCV model.

        Args:
            feature_names: list of feature names. If empty, will return empty
                constraints dictionaries and lists.
            backend: Backend to use for the model. Either "xgboost" or "catboost".
        """
        monotone_constraints = {
            feature_names.index(key): value
            for key, value in self.monotone_constraints.items()
        }

        if backend == "xgboost":
            return {
                "monotone_constraints": monotone_constraints,
                "interaction_constraints": [
                    [feature_names.index(key) for key in lt] for lt in self.interactions
                ],
                "n_jobs": self.n_jobs,
            }
        elif backend == "catboost":
            return {
                "monotone_constraints": monotone_constraints or None,
                "thread_count": self.n_jobs,
            }
        else:
            raise ValueError(
                f"Unknown backend: {backend}. Must be 'xgboost' or 'catboost'."
            )