flexcv.core

flexcv.core.cross_validate(*, X, y, target_name, run, groups, slopes, split_out, split_in, break_cross_val, scale_in, scale_out, n_splits_out, n_splits_in, random_seed, model_effects, mapping, metrics, objective_scorer, em_max_iterations=None, em_stopping_threshold=None, em_stopping_window=None, predict_known_groups_lmm=True, diagnostics=False, **kwargs)

This function performs a cross-validation for a given regression formula, using one or a number of specified machine learning models and a configurable cross-validation method.

Parameters:

Name	Type	Description	Default
X	DataFrame	Features.	required
y	Series	Target.	required
target_name	str	Custom target name.	required
run	Run	A Run object to log to.	required
groups	Series	The grouping or clustering variable.	required
slopes	DataFrame | Series	Random slopes variable(s)	required
split_out	CrossValMethod | BaseCross	Outer split strategy.	required
split_in	CrossValMethod	Inner split strategy.	required
break_cross_val	bool	If True, only the first outer fold is evaluated.	required
scale_in	bool	If True, the features are scaled in the inner cross-validation to zero mean and unit variance. This works independently of the outer scaling.	required
scale_out	bool	If True, the features are scaled in the outer cross-validation to zero mean and unit variance.	required
n_splits_out	int	Number of outer cross-validation folds.	required
n_splits_in	int	Number of inner cross-validation folds.	required
random_seed	int	Seed for all random number generators.	required
model_effects	str	If "fixed", only fixed effects are used. If "mixed", both fixed and random effects are used.	required
mapping	ModelMappingDict	The mapping providing model instances, hyperparameter distributions, and postprocessing functions.	required
metrics	MetricsDict	A dict of metrics to be used as the evaluation metric for the outer cross-validation.	required
objective_scorer	ObjectiveScorer	A custom objective scorer object to provide the evaluation metric for the inner cross-validation.	required
em_max_iterations	int	For use with MERF. Maximum number of iterations for the EM algorithm. (Default: None)	None
em_stopping_threshold	float	For use with MERF. Threshold for the early stopping criterion of the EM algorithm. (Default: None)	None
em_stopping_window	int	For use with MERF. Window size for the early stopping criterion of the EM algorithm. (Default: None)	None
predict_known_groups_lmm	bool	For use with Mixed Linear Models. If True, the model will predict the known groups in the test set. (Default: True)	True
diagnostics	bool	If True, diagnostics plots are logged to Neptune. (Default: False)	False
**kwargs		Additional keyword arguments.	{}

Returns:

Type	Description
Dict[str, Dict[str, list]]	Dict[str, Dict[str, list]]: A dictionary containing the results of the cross-validation, organized by machine learning models.

The function returns a nested dictionary with the following structure:

results_all_folds = {
                    model_name: {
                        "model": [],
                        "parameters": [],
                        "results": [],
                        "r2": [],
                        "y_pred": [],
                        "y_test": [],
                        "shap_values": [],
                        "median_index": [],
                    }

Source code in flexcv/core.py

def cross_validate(
    *,
    X: pd.DataFrame,
    y: pd.Series,
    target_name: str,
    run: NeptuneRun,
    groups: pd.Series,
    slopes: pd.DataFrame | pd.Series,
    split_out: CrossValMethod | BaseCrossValidator | Iterator,
    split_in: CrossValMethod | BaseCrossValidator | Iterator,
    break_cross_val: bool,
    scale_in: bool,
    scale_out: bool,
    n_splits_out: int,
    n_splits_in: int,
    random_seed: int,
    model_effects: str,
    mapping: ModelMappingDict,
    metrics: MetricsDict,
    objective_scorer: ObjectiveScorer,
    em_max_iterations: int = None,
    em_stopping_threshold: float = None,
    em_stopping_window: int = None,
    predict_known_groups_lmm: bool = True,
    diagnostics: bool = False,
    **kwargs,
) -> Dict[str, Dict[str, list]]:
    """This function performs a cross-validation for a given regression formula, using one or a number of specified machine learning models and a configurable cross-validation method.

    Args:
        X (pd.DataFrame): Features.
        y (pd.Series): Target.
        target_name (str): Custom target name.
        run (NeptuneRun): A Run object to log to.
        groups (pd.Series): The grouping or clustering variable.
        slopes (pd.DataFrame | pd.Series): Random slopes variable(s)
        split_out (CrossValMethod | BaseCross): Outer split strategy.
        split_in (CrossValMethod): Inner split strategy.
        break_cross_val (bool): If True, only the first outer fold is evaluated.
        scale_in (bool): If True, the features are scaled in the inner cross-validation to zero mean and unit variance. This works independently of the outer scaling.
        scale_out (bool): If True, the features are scaled in the outer cross-validation to zero mean and unit variance.
        n_splits_out (int): Number of outer cross-validation folds.
        n_splits_in (int): Number of inner cross-validation folds.
        random_seed (int): Seed for all random number generators.
        model_effects (str): If "fixed", only fixed effects are used. If "mixed", both fixed and random effects are used.
        mapping (ModelMappingDict): The mapping providing model instances, hyperparameter distributions, and postprocessing functions.
        metrics (MetricsDict): A dict of metrics to be used as the evaluation metric for the outer cross-validation.
        objective_scorer (ObjectiveScorer): A custom objective scorer object to provide the evaluation metric for the inner cross-validation.
        em_max_iterations (int): For use with MERF. Maximum number of iterations for the EM algorithm. (Default: None)
        em_stopping_threshold (float): For use with MERF. Threshold for the early stopping criterion of the EM algorithm. (Default: None)
        em_stopping_window (int): For use with MERF. Window size for the early stopping criterion of the EM algorithm. (Default: None)
        predict_known_groups_lmm (bool): For use with Mixed Linear Models. If True, the model will predict the known groups in the test set. (Default: True)
        diagnostics (bool): If True, diagnostics plots are logged to Neptune. (Default: False)
        **kwargs: Additional keyword arguments.


    Returns:
      Dict[str, Dict[str, list]]: A dictionary containing the results of the cross-validation, organized by machine learning models.

    The function returns a nested dictionary with the following structure:
    ```python
    results_all_folds = {
                        model_name: {
                            "model": [],
                            "parameters": [],
                            "results": [],
                            "r2": [],
                            "y_pred": [],
                            "y_test": [],
                            "shap_values": [],
                            "median_index": [],
                        }
    ```

    """

    if objective_scorer is None:
        objective_scorer = ObjectiveScorer(mse_wrapper)
    else:
        objective_scorer = ObjectiveScorer(objective_scorer)

    try:
        npt_handler = NeptuneHandler(run=run)
        logger.addHandler(npt_handler)
    except TypeError:
        logger.warning(
            "No Neptune run object passed. Logging to Neptune will be disabled."
        )

    print()
    re_formula = get_re_formula(slopes)
    formula = get_fixed_effects_formula(target_name, X)
    cross_val_split_out = make_cross_val_split(
        method=split_out, groups=groups, n_splits=n_splits_out, random_state=random_seed  # type: ignore
    )

    model_keys = list(mapping.keys())
    for model_name, inner_dict in mapping.items():
        if "add_merf" in inner_dict and inner_dict["add_merf"]:
            merf_name = f"{model_name}_MERF"
            model_keys.append(merf_name)

    results_all_folds = {}

    ######### OUTER FOLD LOOP #########
    for k, (train_index, test_index) in enumerate(
        tqdm(
            cross_val_split_out(X=X, y=y),
            total=n_splits_out,
            desc=" cv",
            position=0,
            leave=False,
        )
    ):
        print()  # for beautiful tqdm progressbar

        groups_exist = groups is not None
        slopes_exist = slopes is not None

        # check if break is requested and if this is the 2nd outer fold
        if break_cross_val and k == 1:
            break

        # Assign the outer folds data
        X_train = X.iloc[train_index]
        y_train = y.iloc[train_index]  # type: ignore

        X_test = X.iloc[test_index]
        y_test = y.iloc[test_index]  # type: ignore

        cluster_train = groups.iloc[train_index] if groups_exist else None  # type: ignore
        cluster_test = groups.iloc[test_index] if groups_exist else None  # type: ignore

        if scale_out:
            X_train_scaled, X_test_scaled = preprocess_features(X_train, X_test)

        else:
            X_train_scaled = X_train
            X_test_scaled = X_test

        if slopes_exist:
            Z_train, Z_test = preprocess_slopes(
                Z_train_slope=slopes.iloc[train_index],
                Z_test_slope=slopes.iloc[test_index],
                must_scale=scale_out,
            )

        else:
            Z_train = np.ones((len(X_train), 1))  # type: ignore
            Z_test = np.ones((len(X_test), 1))  # type: ignore

        # run diagnostics if requested
        if diagnostics:
            to_diagnose = (
                {
                    "effects": model_effects,
                    "cluster_train": cluster_train,
                    "cluster_test": cluster_test,
                }
                if model_effects == "mixed"
                else {"effects": model_effects}
            )

            log_diagnostics(
                X_train_scaled, X_test_scaled, y_train, y_test, run, **to_diagnose
            )

        # Loop over all models
        for model_name in mapping.keys():
            logger.info(f"Evaluating {model_name}...")

            skip_inner_cv = not mapping[model_name]["requires_inner_cv"]

            model_class = mapping[model_name]["model"]

            try:
                model_kwargs = mapping[model_name]["model_kwargs"]
            except KeyError as e:
                raise KeyError(
                    f"No model_kwargs passed for {model_name}. Check your configuration."
                ) from e

            fit_kwargs = mapping[model_name]["fit_kwargs"]
            evaluate_merf = mapping[model_name]["add_merf"]
            param_grid = mapping[model_name]["params"]
            n_trials = mapping[model_name]["n_trials"]
            requires_formula = mapping[model_name]["requires_formula"]

            if requires_formula:
                fit_kwargs["formula"] = formula
                fit_kwargs["re_formula"] = re_formula

            # build inner cv folds
            cross_val_split_in = make_cross_val_split(
                method=split_in, groups=cluster_train, n_splits=n_splits_in, random_state=random_seed  # type: ignore
            )

            if skip_inner_cv:
                # Instantiate model directly without inner cross-validation
                # has to be best_model because it is the only one instantiated
                best_model = model_class(**model_kwargs)
                best_params = best_model.get_params()
                # set study to None since no study is instantiated otherwise
                study = None

            else:
                # this block performs the inner cross-validation with Optuna

                n_trials = mapping[model_name]["n_trials"]
                n_jobs_cv_int = mapping[model_name]["n_jobs_cv"]

                pipe_in = Pipeline(
                    [
                        ("scaler", StandardScaler()) if scale_in else (),
                        (
                            "model",
                            model_class(**model_kwargs),
                        ),
                    ]
                )

                # add "model__" to all keys of the param_grid
                param_grid = add_model_to_keys(param_grid)

                neptune_callback = CustomNeptuneCallback(
                    run[f"{model_name}/Optuna/{k}"],
                    # reduce load on neptune, i.e., reduce number of items and plots
                    log_plot_contour=False,
                    log_plot_optimization_history=True,
                    log_plot_edf=False,
                    study_update_freq=10,  # log every 10th trial,
                )

                # generate numpy random_state object for seeding the sampler
                random_state = check_random_state(random_seed)
                sampler_seed = random_state.randint(0, np.iinfo("int32").max)

                # get sampler to be used for the inner cross-validation
                sampler = optuna.samplers.TPESampler(seed=sampler_seed)

                # instantiate the study object
                study = optuna.create_study(sampler=sampler, direction="maximize")

                # run the inner cross-validation
                study.optimize(
                    lambda trial: objective_cv(
                        trial,
                        cross_val_split=cross_val_split_in,
                        pipe=pipe_in,
                        params=param_grid,
                        X=X_train,
                        y=y_train,
                        run=run,
                        n_jobs=n_jobs_cv_int,
                        objective_scorer=objective_scorer,
                    ),
                    n_jobs=1,
                    n_trials=n_trials,
                    callbacks=[neptune_callback],
                )
                # get best params from study and rm "model__" from keys
                best_params = rm_model_from_keys(study.best_params)

            # add random_state to best_params if it is not already in there
            if "random_state" not in best_params and "random_state" in model_kwargs:
                best_params.update({"random_state": random_seed})

            # add formula to dict if it is required by the model type
            # to_dict can be unpacked in the fit method

            train_pred_kwargs = {}
            test_pred_kwargs = {}
            pred_kwargs = {}

            if mapping[model_name]["consumes_clusters"]:
                fit_kwargs["clusters"] = cluster_train
                fit_kwargs["re_formula"] = re_formula
                pred_kwargs["predict_known_groups_lmm"] = predict_known_groups_lmm

                test_pred_kwargs["clusters"] = cluster_test
                test_pred_kwargs["Z"] = Z_test

                train_pred_kwargs["clusters"] = cluster_train
                train_pred_kwargs["Z"] = Z_train

            # make new instance of the model with the best parameters
            best_model = model_class(
                **handle_duplicate_kwargs(model_kwargs, best_params)
            )

            if "callbacks" not in mapping[model_name]:
                logger.info(f"No callbacks passed for {model_name}. Moving on...")
            elif hasattr(best_model, "callbacks") and hasattr(best_model, "set_params"):
                best_model.set_params(**mapping[model_name]["callbacks"])
            else:
                logger.info(f"Callbacks not supported by {model_name}. Moving on...")

            # Fit the best model on the outer fold
            fit_result = best_model.fit(
                X=X_train_scaled, y=y_train, **handle_duplicate_kwargs(fit_kwargs)
            )
            # get test predictions
            y_pred = best_model.predict(
                X=X_test_scaled,
                **handle_duplicate_kwargs(pred_kwargs, test_pred_kwargs),
            )
            # get training predictions
            y_pred_train = best_model.predict(
                X=X_train_scaled,
                **handle_duplicate_kwargs(pred_kwargs, train_pred_kwargs),
            )

            # store the results of the outer fold of the current model in a dataclass
            # this makes passing to the postprocessor easier
            model_data = SingleModelFoldResult(
                k=k,
                model_name=model_name,
                best_model=best_model,
                best_params=best_params,
                y_pred=y_pred,
                y_test=y_test,
                X_test=X_test_scaled,
                y_pred_train=y_pred_train,
                y_train=y_train,
                X_train=X_train_scaled,
                fit_result=fit_result,
                fit_kwargs=fit_kwargs,
            )
            all_models_dict = model_data.make_results(
                run=run,
                results_all_folds=results_all_folds,
                study=study,
                metrics=metrics,
            )

            try:
                # call model postprocessing on the single results dataclass
                postprocessor = mapping[model_name]["post_processor"]()
                all_models_dict = postprocessor(
                    results_all_folds=results_all_folds,
                    fold_result=model_data,
                    run=run,
                    features=X_train.columns,
                )
            except (KeyError, TypeError):
                logger.info(f"No postprocessor passed for {model_name}. Moving on...")

            if evaluate_merf:
                ###### MERF EVALUATION #################
                # The base model is passed to the MERF class for evaluation in Expectation Maximization (EM) algorithm

                merf_name = f"MERF({model_name})"
                logger.info(f"Evaluating {merf_name}...")

                # tag the base prediction
                y_pred_base = y_pred.copy()

                # instantiate the mixed model with the best fixed effects model
                merf = MERF(
                    fixed_effects_model=mapping[model_name]["model"](**best_params),
                    max_iterations=em_max_iterations,
                    gll_early_stop_threshold=em_stopping_threshold,
                    gll_early_stopping_window=em_stopping_window,
                    log_gll_per_iteration=False,
                )
                # fit the mixed model using cluster variable and Z for slopes
                fit_result = merf.fit(
                    X=X_train_scaled,
                    y=y_train,
                    Z=Z_train,
                    clusters=cluster_train,
                )

                # get test predictions
                y_pred = merf.predict(
                    X=X_test_scaled,
                    clusters=cluster_test,
                    Z=Z_test,
                )
                # get training predictions
                y_pred_train = merf.predict(
                    X=X_train_scaled,
                    clusters=cluster_train,
                    Z=Z_train,
                )

                merf_data = SingleModelFoldResult(
                    k=k,
                    model_name=merf_name,
                    best_model=merf,
                    best_params=best_params,
                    y_pred=y_pred,
                    y_test=y_test,
                    X_test=X_test_scaled,
                    y_pred_train=y_pred_train,
                    y_train=y_train,
                    X_train=X_train_scaled,
                    fit_result=fit_result,
                )

                all_models_dict = merf_data.make_results(
                    run=run,
                    results_all_folds=results_all_folds,
                    study=study,
                    metrics=metrics,
                )

                postprocessor = MERFModelPostProcessor()
                all_models_dict = postprocessor(
                    results_all_folds=all_models_dict,
                    fold_result=merf_data,
                    run=run,
                    y_pred_base=y_pred_base,
                    mixed_name=merf_name,
                )

        print()
        print()

    return results_all_folds

flexcv.core.preprocess_features(X_train, X_test)

Scales the features to zero mean and unit variance.

Parameters:

Name	Type	Description	Default
X_train	DataFrame	Features for the training set.	required
X_test	DataFrame	Features for the test set.	required

Returns:

Type	Description
tuple[DataFrame, DataFrame]	The preprocessed features as a tuple of pandas DataFrames: (X_train_scaled, X_test_scaled)

Source code in flexcv/core.py

def preprocess_features(
    X_train: pd.DataFrame, X_test: pd.DataFrame
) -> tuple[pd.DataFrame, pd.DataFrame]:
    """Scales the features to zero mean and unit variance.

    Args:
        X_train (pd.DataFrame): Features for the training set.
        X_test (pd.DataFrame): Features for the test set.

    Returns:
        (tuple[pd.DataFrame, pd.DataFrame]): The preprocessed features as a tuple of pandas DataFrames: (X_train_scaled, X_test_scaled)
    """
    if not isinstance(X_train, pd.DataFrame):
        raise TypeError(f"X_train must be a pandas DataFrame, not {type(X_train)}")
    if not isinstance(X_test, pd.DataFrame):
        raise TypeError(f"X_test must be a pandas DataFrame, not {type(X_test)}")
    if X_train.shape[1] != X_test.shape[1]:
        raise ValueError(
            f"X_train and X_test must have the same number of columns. X_train has {X_train.shape[1]} columns, X_test has {X_test.shape[1]} columns."
        )

    scaler = StandardScaler()

    X_train_scaled = pd.DataFrame(
        scaler.fit_transform(X_train),
        columns=X_train.columns,
        index=X_train.index,
    )
    X_test_scaled = pd.DataFrame(
        scaler.transform(X_test), columns=X_test.columns, index=X_test.index
    )
    return X_train_scaled, X_test_scaled

flexcv.core.preprocess_slopes(Z_train_slope, Z_test_slope, must_scale)

This function preprocesses the random slopes variable(s) for use in the mixed effects model.

Parameters:

Name	Type	Description	Default
Z_train_slope	DataFrame | Series	Random slopes variable(s) for the training set.	required
Z_test_slope	DataFrame | Series	Random slopes variable(s) for the test set.	required
must_scale	bool	If True, the random slopes are scaled to zero mean and unit variance.	required

Returns:

Type	Description
tuple[ndarray, ndarray]	The preprocessed random slopes as a tuple of numpy arrays: (Z_train, Z_test)

Source code in flexcv/core.py

def preprocess_slopes(
    Z_train_slope: pd.DataFrame | pd.Series,
    Z_test_slope: pd.DataFrame | pd.Series,
    must_scale: bool,
) -> tuple[np.ndarray, np.ndarray]:
    """This function preprocesses the random slopes variable(s) for use in the mixed effects model.

    Args:
        Z_train_slope (pd.DataFrame | pd.Series): Random slopes variable(s) for the training set.
        Z_test_slope (pd.DataFrame | pd.Series): Random slopes variable(s) for the test set.
        must_scale (bool): If True, the random slopes are scaled to zero mean and unit variance.

    Returns:
        (tuple[np.ndarray, np.ndarray]): The preprocessed random slopes as a tuple of numpy arrays: (Z_train, Z_test)
    """
    is_dataframe_train = isinstance(Z_train_slope, pd.DataFrame)
    is_dataframe_test = isinstance(Z_test_slope, pd.DataFrame)
    if not is_dataframe_train and not isinstance(Z_train_slope, pd.Series):
        raise TypeError(
            f"Z_train_slope must be a pandas DataFrame or pandas Series, not {type(Z_train_slope)}"
        )
    if not is_dataframe_test and not isinstance(Z_test_slope, pd.Series):
        raise TypeError(
            f"Z_test_slope must be a pandas DataFrame or pandas Series, not {type(Z_test_slope)}"
        )
    if not isinstance(must_scale, bool):
        raise TypeError(f"must_scale must be a bool, not {type(must_scale)}")

    # check dimensions
    if is_dataframe_train and (Z_train_slope.shape[1] != Z_test_slope.shape[1]):
        raise ValueError(
            f"Z_train_slope and Z_test_slope must have the same number of columns. Z_train_slope has {Z_train_slope.shape[1]} columns, Z_test_slope has {Z_test_slope.shape[1]} columns."
        )
    # convert to DataFrame
    if not is_dataframe_train:
        Z_train_slope = pd.DataFrame(Z_train_slope)
    if not is_dataframe_test:
        Z_test_slope = pd.DataFrame(Z_test_slope)

    if must_scale:
        scaler = StandardScaler()
        Z_train_slope_scaled = pd.DataFrame(
            scaler.fit_transform(Z_train_slope),
            columns=Z_train_slope.columns,
            index=Z_train_slope.index,
        )
        Z_test_slope_scaled = pd.DataFrame(
            scaler.transform(Z_test_slope),
            columns=Z_test_slope.columns,
            index=Z_test_slope.index,
        )
    else:
        Z_train_slope_scaled = Z_train_slope
        Z_test_slope_scaled = Z_test_slope

    Z_train_slope_scaled["Intercept"] = 1
    cols = Z_train_slope_scaled.columns.tolist()
    cols = cols[-1:] + cols[:-1]
    Z_train_slope_scaled = Z_train_slope_scaled[cols]

    Z_test_slope_scaled["Intercept"] = 1
    cols = Z_test_slope_scaled.columns.tolist()
    cols = cols[-1:] + cols[:-1]
    Z_test_slope_scaled = Z_test_slope_scaled[cols]

    Z_train = Z_train_slope_scaled.to_numpy()
    Z_test = Z_test_slope_scaled.to_numpy()
    return Z_train, Z_test