diff --git a/docs/user_guide/selection/MRMR.rst b/docs/user_guide/selection/MRMR.rst index 5fd9a9b18..241c46d93 100644 --- a/docs/user_guide/selection/MRMR.rst +++ b/docs/user_guide/selection/MRMR.rst @@ -509,6 +509,45 @@ which returns the following output: [True, True, False, True, False, True, True, True] +With polars +~~~~~~~~~~~ + +:class:`MRMR()` also works with polars dataframes. It selects the same features and returns +a polars dataframe: + +.. code:: python + + import polars as pl + from sklearn.datasets import load_breast_cancer + from feature_engine.selection import MRMR + + data = load_breast_cancer() + X = pl.DataFrame(data.data, schema=list(data.feature_names), orient="row") + y = pl.Series("target", data.target) + + sel = MRMR(method="FCQ", regression=False) + Xtr = sel.fit_transform(X, y) + print(Xtr.head()) + +In the following output we see the 6 selected features, the same ones that we obtained with +pandas in the F-statistic example: + +.. code:: text + + shape: (5, 6) + ┌────────────────┬──────────────┬──────────────┬─────────────────┬────────────┬───────────────┐ + │ mean perimeter ┆ mean concave ┆ worst radius ┆ worst perimeter ┆ worst area ┆ worst concave │ + │ --- ┆ points ┆ --- ┆ --- ┆ --- ┆ points │ + │ f64 ┆ --- ┆ f64 ┆ f64 ┆ f64 ┆ --- │ + │ ┆ f64 ┆ ┆ ┆ ┆ f64 │ + ╞════════════════╪══════════════╪══════════════╪═════════════════╪════════════╪═══════════════╡ + │ 122.8 ┆ 0.1471 ┆ 25.38 ┆ 184.6 ┆ 2019.0 ┆ 0.2654 │ + │ 132.9 ┆ 0.07017 ┆ 24.99 ┆ 158.8 ┆ 1956.0 ┆ 0.186 │ + │ 130.0 ┆ 0.1279 ┆ 23.57 ┆ 152.5 ┆ 1709.0 ┆ 0.243 │ + │ 77.58 ┆ 0.1052 ┆ 14.91 ┆ 98.87 ┆ 567.7 ┆ 0.2575 │ + │ 135.1 ┆ 0.1043 ┆ 22.54 ┆ 152.2 ┆ 1575.0 ┆ 0.1625 │ + └────────────────┴──────────────┴──────────────┴─────────────────┴────────────┴───────────────┘ + Considerations -------------- diff --git a/feature_engine/selection/base_recursive_selector.py b/feature_engine/selection/base_recursive_selector.py index fe9113077..1161572aa 100644 --- a/feature_engine/selection/base_recursive_selector.py +++ b/feature_engine/selection/base_recursive_selector.py @@ -1,7 +1,9 @@ from types import GeneratorType -from typing import List, Union +from typing import List, Tuple, Union -import pandas as pd +import narwhals as nw +import numpy as np +from narwhals.typing import IntoDataFrame, IntoSeries from sklearn.inspection import permutation_importance from sklearn.model_selection import cross_validate @@ -9,14 +11,13 @@ _check_variables_input_value, ) from feature_engine.dataframe_checks import check_X_y -from feature_engine.selection.base_selection_functions import get_feature_importances +from feature_engine.selection.base_selection_functions import ( + _importance_series, + _select_numerical_variables, + get_feature_importances, +) from feature_engine.selection.base_selector import BaseSelector from feature_engine.tags import _return_tags -from feature_engine.variable_handling import ( - check_numerical_variables, - find_numerical_variables, - retain_variables_if_in_df, -) Variables = Union[None, int, str, List[Union[str, int]]] @@ -81,10 +82,13 @@ class BaseRecursiveSelector(BaseSelector): Performance of the model trained using the original dataset. feature_importances_: - Pandas Series with the feature importance (comes from step 2) + The feature importance (comes from step 2). A pandas Series with the + features as index when X is a pandas dataframe, and a dictionary with the + features as keys otherwise. feature_importances_std_: - Pandas Series with the standard deviation of the feature importance. + The standard deviation of the feature importance, as a pandas Series or a + dictionary, like `feature_importances_`. features_to_drop_: List with the features to remove from the dataset. @@ -116,7 +120,9 @@ def __init__( ): if not isinstance(threshold, (int, float)): - raise ValueError("threshold can only be integer or float") + raise ValueError( + f"threshold must be an integer or a float. Got {threshold} instead." + ) super().__init__(confirm_variables) self.variables = _check_variables_input_value(variables) @@ -126,43 +132,45 @@ def __init__( self.cv = cv self.groups = groups - def fit(self, X: pd.DataFrame, y: pd.Series): + def fit(self, X: IntoDataFrame, y: IntoSeries) -> Tuple[nw.DataFrame, IntoSeries]: """ Find initial model performance. Sort features by importance. Parameters ---------- - X: pandas dataframe of shape = [n_samples, n_features] + X: dataframe of shape = [n_samples, n_features] The input dataframe y: array-like of shape (n_samples) Target variable. Required to train the estimator. - """ - # check input dataframe - X, y = check_X_y(X, y) + Returns + ------- + nw_X: narwhals dataframe + The input dataframe, as a narwhals dataframe. - if self.variables is None: - self.variables_ = find_numerical_variables(X) - else: - if self.confirm_variables is True: - variables_ = retain_variables_if_in_df(X, self.variables) - self.variables_ = check_numerical_variables(X, variables_) - else: - self.variables_ = check_numerical_variables(X, self.variables) + y: Series or numpy array + The target, checked. + """ + nw_X, y = check_X_y(X, y) + + self.variables_ = _select_numerical_variables( + X, self.variables, self.confirm_variables + ) self._cv = list(self.cv) if isinstance(self.cv, GeneratorType) else self.cv # check that there are more than 1 variable to select from self._check_variable_number() - # save input features self._get_feature_names_in(X) + X_model = nw_X.select(nw.col(*self.variables_)).to_native() + # train model with all features and cross-validation model = cross_validate( estimator=self.estimator, - X=X[self.variables_], + X=X_model, y=y, cv=self._cv, groups=self.groups, @@ -170,40 +178,32 @@ def fit(self, X: pd.DataFrame, y: pd.Series): return_estimator=True, ) - # store initial model performance self.initial_model_performance_ = model["test_score"].mean() - # Initialize a dataframe that will contain the list of the feature/coeff - # importance for each cross validation fold - feature_importances_cv = pd.DataFrame() - - # Populate the feature_importances_cv dataframe with columns containing - # the feature importance values for each model returned by the cross - # validation. - # There are as many columns as folds. - for i in range(len(model["estimator"])): - m = model["estimator"][i] - + # one row of feature importance per cross-validation fold + importances = [] + for m in model["estimator"]: if hasattr(m, "feature_importances_") or hasattr(m, "coef_"): - feature_importances_cv[i] = get_feature_importances(m) + importances.append(get_feature_importances(m)) else: r = permutation_importance( m, - X[self.variables_], + X_model, y, n_repeats=1, random_state=10, ) - feature_importances_cv[i] = r.importances_mean + importances.append(r.importances_mean) + importances_arr = np.array(importances) - # Add the variables as index to feature_importances_cv - feature_importances_cv.index = self.variables_ - - # Aggregate the feature importance returned in each fold - self.feature_importances_ = feature_importances_cv.mean(axis=1) - self.feature_importances_std_ = feature_importances_cv.std(axis=1) + self.feature_importances_ = _importance_series( + X, self.variables_, importances_arr.mean(axis=0) + ) + self.feature_importances_std_ = _importance_series( + X, self.variables_, importances_arr.std(axis=0, ddof=1) + ) - return X, y + return nw_X, y def _more_tags(self): tags_dict = _return_tags() diff --git a/feature_engine/selection/base_selection_functions.py b/feature_engine/selection/base_selection_functions.py index 96fc45970..4cbadb60b 100644 --- a/feature_engine/selection/base_selection_functions.py +++ b/feature_engine/selection/base_selection_functions.py @@ -1,8 +1,11 @@ -from typing import List, Union from types import GeneratorType +from typing import List, Union +import narwhals as nw +import narwhals.dependencies as nwd import numpy as np -import pandas as pd +from narwhals.typing import IntoDataFrame +from scipy.stats import kendalltau, rankdata from sklearn.model_selection import cross_validate from feature_engine.variable_handling import ( @@ -36,8 +39,19 @@ def get_feature_importances(estimator): return importances +def _importance_series(X: IntoDataFrame, features, values: np.ndarray): + """ + Return the importance of each feature as a pandas Series indexed by the + features when X is a pandas dataframe, or as a dictionary with the features as + keys otherwise. + """ + if nwd.is_pandas_dataframe(X) is True: + return nw.get_native_namespace(X).Series(values, index=features) + return dict(zip(features, values.tolist())) + + def _select_all_variables( - X: pd.DataFrame, + X: IntoDataFrame, variables: Variables, confirm_variables: bool, exclude_datetime: bool = False, @@ -62,7 +76,7 @@ def _select_all_variables( def _select_numerical_variables( - X: pd.DataFrame, + X: IntoDataFrame, variables: Variables, confirm_variables: bool, ): @@ -85,8 +99,113 @@ def _select_numerical_variables( return variables_ +def _corrcoef(values: np.ndarray) -> np.ndarray: + # constant columns return NaN, like pandas, instead of warning. + with np.errstate(divide="ignore", invalid="ignore"): + return np.corrcoef(values, rowvar=False) + + +def _pearson_pairwise_complete(values: np.ndarray, finite: np.ndarray) -> np.ndarray: + """ + Pearson correlation of every pair of columns, using the rows where both are + finite. The sums of all pairs come from matrix products, which is much faster + than looping over the pairs. + """ + mask: np.ndarray = finite.astype(float) + with np.errstate(divide="ignore", invalid="ignore"): + # centring first keeps the sums below numerically stable. + mean = np.where(finite, values, 0.0).sum(axis=0) / mask.sum(axis=0) + x = np.where(finite, values - mean, 0.0) + n_obs = mask.T @ mask + sum_x = x.T @ mask + sum_xx = (x * x).T @ mask + var = sum_xx - sum_x * sum_x / n_obs + corr = (x.T @ x - sum_x * sum_x.T / n_obs) / np.sqrt(var * var.T) + + # when the variance of the shared rows is tiny compared with the sums, the + # subtraction above loses precision: recompute those pairs one by one. + unstable = (var <= 1e-8 * sum_xx) | (var.T <= 1e-8 * sum_xx.T) + corr[unstable] = np.nan + for i, j in zip(*np.nonzero(np.triu(unstable & (n_obs > 1), 1))): + rows = finite[:, i] & finite[:, j] + corr[i, j] = _corrcoef(x[rows][:, [i, j]])[0, 1] + return corr + + +def _spearman_pairwise_complete(nw_X: nw.DataFrame) -> np.ndarray: + """ + Spearman correlation of every pair of columns, ranking each pair on the rows + where both are finite, like pandas.DataFrame.corr(). + """ + variables = nw_X.columns + exprs = [] + for i, var_i in enumerate(variables): + for j in range(i + 1, len(variables)): + var_j = variables[j] + rows = nw.col(var_i).is_finite() & nw.col(var_j).is_finite() + x = nw.when(rows).then(nw.col(var_i)).rank("average") + y = nw.when(rows).then(nw.col(var_j)).rank("average") + dx = x - x.mean() + dy = y - y.mean() + exprs.append( + ((dx * dy).sum() / ((dx * dx).sum() * (dy * dy).sum()).sqrt()).alias( + f"__{i}_{j}__" + ) + ) + n_vars = len(variables) + corr = np.full((n_vars, n_vars), np.nan) + corr[np.triu_indices(n_vars, 1)] = np.array(nw_X.select(exprs).row(0), dtype=float) + return corr + + +def _correlation_matrix(X: IntoDataFrame, variables: list, method) -> np.ndarray: + """ + Correlation matrix of the variables. Like pandas.DataFrame.corr(), each pair of + variables is compared on the rows where both have finite values. Only the + values above the diagonal are used. + """ + if nwd.is_pandas_dataframe(X) is True: + values = X[variables].to_numpy(dtype=float, na_value=np.nan) + else: + nw_X = nw.from_native(X, eager_only=True).select(nw.col(*variables)) + values = nw_X.to_numpy().astype(float) + finite = np.isfinite(values) + + # numpy is faster than pandas and narwhals. + if method == "pearson": + if finite.all(): + return _corrcoef(values) + return _pearson_pairwise_complete(values, finite) + + if method == "spearman" and finite.all(): + # scipy ranks faster than pandas, and polars faster than scipy. + if nwd.is_pandas_dataframe(X) is True: + return _corrcoef(rankdata(values, axis=0)) + return _corrcoef(nw_X.select(nw.all().rank("average")).to_numpy()) + + # pandas is faster than narwhals. + if nwd.is_pandas_dataframe(X) is True: + return X[variables].corr(method=method).to_numpy() + + if method == "spearman": + return _spearman_pairwise_complete(nw_X) + + # kendall and callables are computed pair by pair, as pandas does. + corr_func = (lambda a, b: kendalltau(a, b)[0]) if method == "kendall" else method + n_vars = len(variables) + corr = np.full((n_vars, n_vars), np.nan) + for i in range(n_vars): + for j in range(i + 1, n_vars): + rows = finite[:, i] & finite[:, j] + if rows.all(): + corr[i, j] = corr_func(values[:, i], values[:, j]) + elif rows.any(): + corr[i, j] = corr_func(values[rows, i], values[rows, j]) + return corr + + def find_correlated_features( - X: pd.DataFrame, + X: IntoDataFrame, variables: list[Union[str, int]], method: str, threshold: float, @@ -96,7 +215,7 @@ def find_correlated_features( Parameters ---------- - X : pandas dataframe of shape = [n_samples, n_features] + X : dataframe of shape = [n_samples, n_features] The training dataset. variables : list @@ -133,36 +252,32 @@ def find_correlated_features( correlated with the key. The key + the values should be the same as the set found in `correlated_feature_groups`. """ - # the correlation matrix - correlated_matrix = X[variables].corr(method=method).to_numpy() + correlated_matrix = _correlation_matrix(X, variables, method) # the correlated pairs correlated_mask = np.triu(np.abs(correlated_matrix), 1) > threshold - examined = set() + examined: np.ndarray = np.zeros(len(variables), dtype=bool) correlated_groups = list() features_to_drop = list() correlated_dict = {} for i, f_i in enumerate(variables): - if f_i not in examined: - examined.add(f_i) - temp_set = set([f_i]) - for j, f_j in enumerate(variables): - if f_j not in examined: - if correlated_mask[i, j] == 1: - examined.add(f_j) - features_to_drop.append(f_j) - temp_set.add(f_j) - if len(temp_set) > 1: - correlated_groups.append(temp_set) - correlated_dict[f_i] = temp_set.difference({f_i}) + if examined.item(i) is False: + examined[i] = True + correlated = np.flatnonzero(correlated_mask[i] & ~examined) + if len(correlated) > 0: + examined[correlated] = True + correlated_features = [variables[j] for j in correlated] + features_to_drop.extend(correlated_features) + correlated_groups.append({f_i, *correlated_features}) + correlated_dict[f_i] = set(correlated_features) return correlated_groups, features_to_drop, correlated_dict def single_feature_performance( - X: pd.DataFrame, - y: pd.Series, + X: IntoDataFrame, + y, variables: List[Union[str, int]], estimator, cv, @@ -174,7 +289,7 @@ def single_feature_performance( Parameters ---------- - X: pandas dataframe of shape = [n_samples, n_features] + X: dataframe of shape = [n_samples, n_features] The input dataframe y: array-like of shape (n_samples) @@ -211,12 +326,13 @@ def single_feature_performance( feature_performance_std = {} cv = list(cv) if isinstance(cv, GeneratorType) else cv + nw_X = nw.from_native(X, eager_only=True) # train a model for every feature and store the performance for feature in variables: model = cross_validate( estimator, - X[feature].to_frame(), + nw_X.get_column(feature).to_frame().to_native(), y, cv=cv, groups=groups, @@ -230,8 +346,8 @@ def single_feature_performance( def find_feature_importance( - X: pd.DataFrame, - y: pd.Series, + X: IntoDataFrame, + y, estimator, cv, scoring, @@ -245,7 +361,7 @@ def find_feature_importance( Parameters ---------- - X: pandas dataframe of shape = [n_samples, n_features] + X: dataframe of shape = [n_samples, n_features] The input dataframe y: array-like of shape (n_samples) @@ -268,14 +384,15 @@ def find_feature_importance( Returns ------- - feature_importance: pd.Series - A pandas Series with the feature name as index and its importance as value. The - importance is given by the coefficients of linear models or the impurity gain - from tree-based models. - - feature_importance_std: pd.Series - A pandas Series with the feature name as key and the standard deviation of the - feature importance as value. + feature_importance: pandas Series or dict + The importance of each feature, given by the coefficients of linear models or + the impurity gain from tree-based models. A pandas Series with the feature + names as index when X is a pandas dataframe, and a dictionary with the + feature names as keys otherwise. + + feature_importance_std: pandas Series or dict + The standard deviation of the importance of each feature, as a pandas Series + or a dictionary, like `feature_importance`. """ cv = list(cv) if isinstance(cv, GeneratorType) else cv @@ -289,19 +406,16 @@ def find_feature_importance( return_estimator=True, ) - # dataframe to store the feature importance for each cv fold - feature_importances_cv = pd.DataFrame() - - # Populate dataframe with columns containing the feature importance values - # for each cv fold. There are as many columns as folds. - for i in range(len(model["estimator"])): - m = model["estimator"][i] - feature_importances_cv[i] = get_feature_importances(m) + importances = np.array([get_feature_importances(m) for m in model["estimator"]]) - # add the variables as the index to feature_importances_cv - feature_importances_cv.index = X.columns + # pandas keeps the columns index, with its name and dtype. + if nwd.is_pandas_dataframe(X) is True: + features = X.columns + else: + features = nw.from_native(X, eager_only=True).columns - # aggregate the feature importance returned in each fold - feature_importances_ = feature_importances_cv.mean(axis=1) - feature_importances_std_ = feature_importances_cv.std(axis=1) + feature_importances_ = _importance_series(X, features, importances.mean(axis=0)) + feature_importances_std_ = _importance_series( + X, features, importances.std(axis=0, ddof=1) + ) return feature_importances_, feature_importances_std_ diff --git a/feature_engine/selection/base_selector.py b/feature_engine/selection/base_selector.py index cfa8f1c95..0e4ff5d97 100644 --- a/feature_engine/selection/base_selector.py +++ b/feature_engine/selection/base_selector.py @@ -1,5 +1,7 @@ +import narwhals as nw +import narwhals.dependencies as nwd import numpy as np -import pandas as pd +from narwhals.typing import IntoDataFrame from sklearn.base import BaseEstimator, TransformerMixin from sklearn.utils.validation import check_is_fitted @@ -41,41 +43,43 @@ def __init__( self.confirm_variables = confirm_variables - def transform(self, X: pd.DataFrame) -> pd.DataFrame: + def transform(self, X: IntoDataFrame) -> IntoDataFrame: """ Return dataframe with selected features. Parameters ---------- - X: pandas dataframe of shape = [n_samples, n_features]. + X: dataframe of shape = [n_samples, n_features]. The input dataframe. Returns ------- - X_new: pandas dataframe of shape = [n_samples, n_selected_features] - Pandas dataframe with the selected features. + X_new: dataframe of shape = [n_samples, n_selected_features] + The dataframe with the selected features, in the same library as the + input. """ - - # check if fit is performed prior to transform check_is_fitted(self) - - # check if input is a dataframe - X = check_X(X) - - # check if number of columns in test dataset matches to train dataset + nw_X = check_X(X) _check_X_matches_training_df(X, self.n_features_in_) - # reorder df to match train set - X = X[self.feature_names_in_] + # selecting in the train set order also restores the train column order. + features_to_drop = set(self.features_to_drop_) + features = [f for f in self.feature_names_in_ if f not in features_to_drop] - # return the dataframe with the selected features - return X.drop(columns=self.features_to_drop_) + # pandas is faster than narwhals. + if nwd.is_pandas_dataframe(X) is True: + return X[features] + else: + return nw_X.select(nw.col(*features)).to_native() - def _get_feature_names_in(self, X): - """Get the names and number of features in the train set. The dataframe - used during fit.""" + def _get_feature_names_in(self, X: IntoDataFrame): + """Get the names and number of features in the train set (the dataframe + used during fit).""" - self.feature_names_in_ = X.columns.to_list() + if nwd.is_pandas_dataframe(X) is True: + self.feature_names_in_ = list(X.columns) + else: + self.feature_names_in_ = nw.from_native(X, eager_only=True).columns self.n_features_in_ = X.shape[1] return self diff --git a/feature_engine/selection/mrmr.py b/feature_engine/selection/mrmr.py index f5d2c9f9a..7683044c4 100644 --- a/feature_engine/selection/mrmr.py +++ b/feature_engine/selection/mrmr.py @@ -1,10 +1,10 @@ -import copy - from types import GeneratorType from typing import List, Optional, Union +import narwhals as nw +import narwhals.dependencies as nwd import numpy as np -import pandas as pd +from narwhals.typing import IntoDataFrame, IntoSeries from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor from sklearn.feature_selection import ( f_classif, @@ -40,13 +40,12 @@ _CLASSIFICATION_METRICS, _REGRESSION_METRICS, ) +from feature_engine.selection.base_selection_functions import ( + _correlation_matrix, + _select_numerical_variables, +) from feature_engine.selection.base_selector import BaseSelector from feature_engine.tags import _return_tags -from feature_engine.variable_handling import ( - check_numerical_variables, - find_numerical_variables, - retain_variables_if_in_df, -) _cv_docstring = _cv_docstring + """ Only used when `method = 'RFCQ'`.""" @@ -153,7 +152,7 @@ class MRMR(BaseSelector): will optimize the 'max_depth' over `[1, 2, 3, 4]`. Only used when `method` is `'RFCQ'`. - regression: boolean, default=True + regression: boolean, default=False Indicates whether the target is one for regression or a classification. {confirm_variables} @@ -216,6 +215,31 @@ class MRMR(BaseSelector): 2 7.2574 2.802260 3 5.6431 2.547945 4 3.8462 2.181467 + + With polars + + >>> import polars as pl + >>> from sklearn.datasets import fetch_california_housing + >>> from feature_engine.selection import MRMR + >>> data = fetch_california_housing() + >>> X = pl.DataFrame(data.data, schema=data.feature_names, orient="row") + >>> X = X.drop("Latitude", "Longitude") + >>> y = pl.Series("MedHouseVal", data.target) + >>> mrmr_sel = MRMR(method="MIQ", regression=True, random_state=3) + >>> X_t = mrmr_sel.fit_transform(X, y) + >>> print(X_t.head()) + shape: (5, 2) + ┌────────┬──────────┐ + │ MedInc ┆ AveOccup │ + │ --- ┆ --- │ + │ f64 ┆ f64 │ + ╞════════╪══════════╡ + │ 8.3252 ┆ 2.555556 │ + │ 8.3014 ┆ 2.109842 │ + │ 7.2574 ┆ 2.80226 │ + │ 5.6431 ┆ 2.547945 │ + │ 3.8462 ┆ 2.181467 │ + └────────┴──────────┘ """ def __init__( @@ -284,7 +308,7 @@ def __init__( and scoring not in _CLASSIFICATION_METRICS ): raise ValueError( - f"The metric {scoring} is not suitable for classification. Set the" + f"The metric {scoring} is not suitable for classification. Set the " "parameter regression to True or choose a different performance " "metric." ) @@ -302,88 +326,97 @@ def __init__( self.random_state = random_state self.n_jobs = n_jobs - def fit(self, X: pd.DataFrame, y: pd.Series): + def fit(self, X: IntoDataFrame, y: IntoSeries): """ Find the important features. Parameters ---------- - X: pandas dataframe of shape = [n_samples, n_features] + X: dataframe of shape = [n_samples, n_features] The input dataframe. y: array-like of shape (n_samples) Target variable. Required to train the estimator. """ - # check input dataframe - X, y = check_X_y(X, y) + nw_X, y = check_X_y(X, y) - if self.variables is None: - self.variables_ = find_numerical_variables(X) - else: - if self.confirm_variables is True: - variables_ = retain_variables_if_in_df(X, self.variables) - self.variables_ = check_numerical_variables(X, variables_) - else: - self.variables_ = check_numerical_variables(X, self.variables) + self.variables_ = _select_numerical_variables( + X, self.variables, self.confirm_variables + ) - # check that there are more than 1 variable to select from self._check_variable_number() - - # save input features self._get_feature_names_in(X) - # start the search - self.relevance_ = self._calculate_relevance(X[self.variables_], y) - - # select most relevant feature - relevance = self.relevance_ - n = relevance.argmax() # most relevant feature - relevance = np.delete(relevance, n) + # pandas is faster than narwhals. + if nwd.is_pandas_dataframe(X) is True: + X_vars = X[self.variables_] + else: + nw_X_vars = nw_X.select(nw.col(*self.variables_)) + X_vars = nw_X_vars.to_native() - remaining = copy.deepcopy(self.variables_) # all variables to examine - feature = remaining[n] # most important feature so far - selected = [feature] # selected feature so far - remaining.remove(feature) # remaining features to examine + # scikit-learn's f_regression fails with polars series. + if nwd.is_into_series(y) is True and nwd.is_pandas_series(y) is False: + y = nw.from_native(y, series_only=True).to_numpy() - # calculate redundance - redundance = self._calculate_redundance(X[remaining], X[feature]) + self.relevance_ = self._calculate_relevance(X_vars, y) - # find feature with highest mrmr - mrmr = self._calculate_mrmr(relevance, redundance) + if self.method in ["FCD", "FCQ", "RFCQ"]: + # one matrix product for all pairs is much faster than correlating the + # remaining features at each iteration. + corr = np.triu(_correlation_matrix(X, self.variables_, "pearson"), 1) + corr = np.absolute(corr + corr.T) + else: + # pandas is faster than narwhals. + if nwd.is_pandas_dataframe(X) is True: + values = X_vars.to_numpy(dtype=float, na_value=np.nan) + else: + values = nw_X_vars.to_numpy() if self.max_features is None: - iter = int(0.2 * len(self.variables_)) - 2 + n_features = int(0.2 * len(self.variables_)) else: - iter = self.max_features - 2 + n_features = self.max_features - for i in range(iter): - # find feature with maximum relevance and minimum redundancy - n = mrmr.argmax() + remaining: np.ndarray = np.ones(len(self.variables_), dtype=bool) + selected = [self.relevance_.argmax()] + redundance_sum = np.zeros(len(self.variables_)) - # adjust feature lists - feature = remaining[n] - selected.append(feature) - remaining.remove(feature) + # at least 2 features are selected, also when max_features is 1. + while len(selected) < max(n_features, 2): + feature = selected[-1] + remaining[feature] = False + to_examine = np.flatnonzero(remaining) - relevance = np.delete(relevance, n) - if i == 0: - redundance = np.delete(redundance, n) + if self.method in ["FCD", "FCQ", "RFCQ"]: + redundance = corr[to_examine, feature] else: - redundance = np.delete(redundance, n, axis=1) - - new_redundance = self._calculate_redundance(X[remaining], X[feature]) - redundance = np.vstack([redundance, new_redundance]) - mean_redundance = redundance.mean(axis=0) - - mrmr = self._calculate_mrmr(relevance, mean_redundance) + redundance = self._mutual_information( + values[:, to_examine], self._feature_values(X, nw_X, feature) + ) - n = mrmr.argmax() - selected.append(remaining[n]) + # summing row by row and dividing gives the same values as the mean + # of the redundance to all the selected features. + redundance_sum[to_examine] += redundance + mrmr = self._calculate_mrmr( + self.relevance_[to_examine], + redundance_sum[to_examine] / len(selected), + ) + selected.append(to_examine[mrmr.argmax()]) - self.features_to_drop_ = [f for f in self.variables_ if f not in selected] + selected_features = {self.variables_[i] for i in selected} + self.features_to_drop_ = [ + f for f in self.variables_ if f not in selected_features + ] return self + def _feature_values(self, X: IntoDataFrame, nw_X: nw.DataFrame, i) -> np.ndarray: + """Values of the i-th variable, in its own dtype.""" + # pandas is faster than narwhals. + if nwd.is_pandas_dataframe(X) is True: + return X[self.variables_[i]].to_numpy() + return nw_X.get_column(self.variables_[i]).to_numpy() + def _calculate_relevance(self, X, y): if self.method in ["MIQ", "MID"]: @@ -424,10 +457,10 @@ def _calculate_relevance(self, X, y): n_jobs=self.n_jobs, ) - if self.param_grid: - param_grid = self.param_grid - else: + if self.param_grid is None or len(self.param_grid) == 0: param_grid = {"max_depth": [1, 2, 3, 4]} + else: + param_grid = self.param_grid cv = list(self.cv) if isinstance(self.cv, GeneratorType) else self.cv @@ -441,22 +474,17 @@ def _calculate_relevance(self, X, y): return relevance - def _calculate_redundance(self, X, y): - - if self.method in ["FCD", "FCQ", "RFCQ"]: - redundance = X.corrwith(y).values - redundance = np.absolute(redundance) - - else: - redundance = mutual_info_regression( - X=X, - y=y, - n_neighbors=self.n_neighbors, - random_state=self.random_state, - n_jobs=self.n_jobs, - ) - - return redundance + def _mutual_information(self, X: np.ndarray, y: np.ndarray) -> np.ndarray: + """Mutual information between each column of X and the feature y.""" + # X is a copy made by the caller, so scikit-learn doesn't need to copy it. + return mutual_info_regression( + X=X, + y=y, + n_neighbors=self.n_neighbors, + copy=False, + random_state=self.random_state, + n_jobs=self.n_jobs, + ) def _calculate_mrmr(self, relevance, redundance): if self.method in ["MID", "FCD"]: diff --git a/tests/test_selection/conftest.py b/tests/test_selection/conftest.py index e41d7ce4e..7006979c8 100644 --- a/tests/test_selection/conftest.py +++ b/tests/test_selection/conftest.py @@ -25,6 +25,23 @@ def df_test(): return X, y +@pytest.fixture(scope="module") +def data_classification(): + """The data of df_test as a plain dict, with the target under "target".""" + X, y = make_classification( + n_samples=1000, + n_features=12, + n_redundant=4, + n_clusters_per_class=1, + weights=[0.50], + class_sep=2, + random_state=1, + ) + data = {f"var_{i}": X[:, i].tolist() for i in range(12)} + data["target"] = y.tolist() + return data + + @pytest.fixture(scope="module") def df_test_with_groups(): # Parameters diff --git a/tests/test_selection/test_base_recursive_selector.py b/tests/test_selection/test_base_recursive_selector.py new file mode 100644 index 000000000..a4b0a3c6b --- /dev/null +++ b/tests/test_selection/test_base_recursive_selector.py @@ -0,0 +1,257 @@ +import re + +import narwhals as nw +import numpy as np +import pandas as pd +import pytest +from sklearn.ensemble import RandomForestClassifier +from sklearn.linear_model import LogisticRegression +from sklearn.model_selection import GroupKFold, StratifiedKFold +from sklearn.neighbors import KNeighborsClassifier + +from feature_engine.selection.base_recursive_selector import BaseRecursiveSelector +from tests.backend_helpers import make_series + +VARIABLES = ["var_0", "var_4", "var_7"] + + +def _split_target(make_df, data): + X = make_df({k: v for k, v in data.items() if k != "target"}) + return X, make_series(make_df, data["target"]) + + +# init parameters +@pytest.mark.parametrize("threshold", [None, [0.1], "a_string", {"a": 1}]) +def test_error_if_threshold_not_number(threshold): + msg = f"threshold must be an integer or a float. Got {threshold} instead." + with pytest.raises(ValueError, match=re.escape(msg)): + BaseRecursiveSelector(RandomForestClassifier(), threshold=threshold) + + +@pytest.mark.parametrize( + "estimator, scoring, cv, groups, threshold, confirm_variables", + [ + (RandomForestClassifier(), "roc_auc", 3, None, 0.01, False), + (LogisticRegression(), "accuracy", StratifiedKFold(), [1, 2], 1, True), + (KNeighborsClassifier(), "r2", GroupKFold(), None, -0.5, False), + ], +) +def test_init_param_assignment( + estimator, scoring, cv, groups, threshold, confirm_variables +): + selector = BaseRecursiveSelector( + estimator, + scoring=scoring, + cv=cv, + groups=groups, + threshold=threshold, + confirm_variables=confirm_variables, + ) + assert selector.estimator is estimator + assert selector.scoring == scoring + assert selector.cv is cv + assert selector.groups == groups + assert selector.threshold == threshold + assert selector.confirm_variables is confirm_variables + + +# fit +@pytest.mark.parametrize("cv_type", ["int", "splitter", "generator"]) +def test_fit_with_feature_importances(make_df, data_classification, cv_type): + X, y = _split_target(make_df, data_classification) + cv = {"int": 3, "splitter": StratifiedKFold(n_splits=3)}.get(cv_type) + if cv_type == "generator": + cv = StratifiedKFold(n_splits=3).split(X, y) + + selector = BaseRecursiveSelector( + RandomForestClassifier(n_estimators=5, random_state=1), + scoring="roc_auc", + cv=cv, + variables=VARIABLES, + ) + selector.fit(X, y) + + assert selector.variables_ == VARIABLES + assert selector.feature_names_in_ == [f"var_{i}" for i in range(12)] + assert selector.n_features_in_ == 12 + assert selector.initial_model_performance_ == pytest.approx(0.9947395643178775) + assert dict(selector.feature_importances_) == pytest.approx( + { + "var_0": 0.05016049596989658, + "var_4": 0.5704427408209541, + "var_7": 0.37939676320914945, + } + ) + assert dict(selector.feature_importances_std_) == pytest.approx( + { + "var_0": 0.02070511883333705, + "var_4": 0.031350384984021235, + "var_7": 0.03942210400299374, + } + ) + + +def test_fit_with_coefficients(make_df, data_classification): + X, y = _split_target(make_df, data_classification) + selector = BaseRecursiveSelector( + LogisticRegression(), scoring="roc_auc", cv=3, variables=VARIABLES + ) + selector.fit(X, y) + + assert selector.initial_model_performance_ == pytest.approx(0.996732746280939) + assert dict(selector.feature_importances_) == pytest.approx( + { + "var_0": 2.0421118575507067, + "var_4": 0.36861802531867144, + "var_7": 2.68269483714549, + } + ) + assert dict(selector.feature_importances_std_) == pytest.approx( + { + "var_0": 0.08627973281236784, + "var_4": 0.12490483002792199, + "var_7": 0.13862536091514688, + } + ) + + +def test_fit_with_permutation_importance(make_df, data_classification): + # KNN has no coef_ or feature_importances_, so importance comes from + # permutation_importance. + X, y = _split_target(make_df, data_classification) + selector = BaseRecursiveSelector( + KNeighborsClassifier(), scoring="accuracy", cv=3, variables=VARIABLES + ) + selector.fit(X, y) + + assert selector.initial_model_performance_ == pytest.approx(0.991997986009962) + assert dict(selector.feature_importances_) == pytest.approx( + { + "var_0": 0.0050000000000000044, + "var_4": 0.0753333333333333, + "var_7": 0.42766666666666664, + } + ) + assert dict(selector.feature_importances_std_) == pytest.approx( + { + "var_0": 0.0010000000000000009, + "var_4": 0.0005773502691896263, + "var_7": 0.0037859388972001223, + } + ) + + +def test_feature_importances_type(make_df, data_classification): + X, y = _split_target(make_df, data_classification) + selector = BaseRecursiveSelector( + LogisticRegression(), scoring="roc_auc", cv=3, variables=VARIABLES + ) + selector.fit(X, y) + + importance_type = pd.Series if make_df is pd.DataFrame else dict + assert isinstance(selector.feature_importances_, importance_type) + assert isinstance(selector.feature_importances_std_, importance_type) + assert list(selector.feature_importances_.keys()) == VARIABLES + + +def test_fit_returns_narwhals_frame_and_target(make_df, data_classification): + X, y = _split_target(make_df, data_classification) + selector = BaseRecursiveSelector( + LogisticRegression(), scoring="roc_auc", cv=3, variables=VARIABLES + ) + nw_X, y_ = selector.fit(X, y) + + assert isinstance(nw_X, nw.DataFrame) + assert nw_X.to_native() is X + assert isinstance(y_, type(y)) + + +def test_fit_finds_numerical_variables(make_df, data_classification): + X, y = _split_target( + make_df, + { + "var_0": data_classification["var_0"], + "cat": ["a", "b"] * 500, + "var_4": data_classification["var_4"], + "target": data_classification["target"], + }, + ) + selector = BaseRecursiveSelector(LogisticRegression(), scoring="roc_auc", cv=3) + selector.fit(X, y) + + assert selector.variables_ == ["var_0", "var_4"] + assert selector.feature_names_in_ == ["var_0", "cat", "var_4"] + assert list(selector.feature_importances_.keys()) == ["var_0", "var_4"] + + +def test_fit_with_confirm_variables(make_df, data_classification): + X, y = _split_target(make_df, data_classification) + selector = BaseRecursiveSelector( + LogisticRegression(), + scoring="roc_auc", + cv=3, + variables=["var_0", "var_4", "Hola"], + confirm_variables=True, + ) + selector.fit(X, y) + + assert selector.variables_ == ["var_0", "var_4"] + assert list(selector.feature_importances_.keys()) == ["var_0", "var_4"] + + +@pytest.mark.parametrize("target_type", [list, np.array]) +def test_fit_with_list_and_array_target(make_df, data_classification, target_type): + X, _ = _split_target(make_df, data_classification) + y = target_type(data_classification["target"]) + selector = BaseRecursiveSelector( + LogisticRegression(), scoring="roc_auc", cv=3, variables=VARIABLES + ) + selector.fit(X, y) + + assert selector.initial_model_performance_ == pytest.approx(0.996732746280939) + + +def test_error_if_only_one_variable(make_df, data_classification): + X, y = _split_target(make_df, data_classification) + selector = BaseRecursiveSelector( + LogisticRegression(), scoring="roc_auc", cv=3, variables=["var_0"] + ) + msg = ( + "The selector needs at least 2 or more variables to select from. " + "Got only 1 variable: ['var_0']." + ) + with pytest.raises(ValueError, match=re.escape(msg)): + selector.fit(X, y) + + +def test_feature_importances_are_pandas_series_indexed_by_variables( + data_classification, +): + X, y = _split_target(pd.DataFrame, data_classification) + selector = BaseRecursiveSelector( + LogisticRegression(), scoring="roc_auc", cv=3, variables=VARIABLES + ) + selector.fit(X, y) + + pd.testing.assert_series_equal( + selector.feature_importances_, + pd.Series( + [2.0421118575507067, 0.36861802531867144, 2.68269483714549], + index=VARIABLES, + ), + ) + + +def test_fit_with_integer_column_names(data_classification): + X, y = _split_target(pd.DataFrame, data_classification) + X.columns = list(range(12)) + selector = BaseRecursiveSelector( + LogisticRegression(), scoring="roc_auc", cv=3, variables=[0, 4, 7] + ) + selector.fit(X, y) + + assert selector.variables_ == [0, 4, 7] + assert selector.feature_names_in_ == list(range(12)) + assert dict(selector.feature_importances_) == pytest.approx( + {0: 2.0421118575507067, 4: 0.36861802531867144, 7: 2.68269483714549} + ) diff --git a/tests/test_selection/test_base_selection_functions.py b/tests/test_selection/test_base_selection_functions.py index b2345a53e..0c4cb0e03 100644 --- a/tests/test_selection/test_base_selection_functions.py +++ b/tests/test_selection/test_base_selection_functions.py @@ -1,333 +1,364 @@ +from datetime import datetime + +import numpy as np import pandas as pd import pytest from sklearn.ensemble import RandomForestClassifier -from sklearn.model_selection import StratifiedKFold, GroupKFold - +from sklearn.linear_model import Lasso, LogisticRegression +from sklearn.model_selection import GroupKFold, StratifiedKFold from feature_engine.selection.base_selection_functions import ( _select_all_variables, _select_numerical_variables, find_correlated_features, find_feature_importance, + get_feature_importances, single_feature_performance, ) - - -@pytest.fixture -def df(): - df = pd.DataFrame( - { - "Name": ["tom", "nick", "krish", "jack"], - "City": ["London", "Manchester", "Liverpool", "Bristol"], - "Age": [20, 21, 19, 18], - "Marks": [0.9, 0.8, 0.7, 0.6], - "date_range": pd.date_range("2020-02-24", periods=4, freq="min"), - "date_obj0": ["2020-02-24", "2020-02-25", "2020-02-26", "2020-02-27"], - } +from tests.backend_helpers import make_series + +DATA_VARTYPES = { + "Name": ["tom", "nick", "krish", "jack"], + "City": ["London", "Manchester", "Liverpool", "Bristol"], + "Age": [20, 21, 19, 18], + "Marks": [0.9, 0.8, 0.7, 0.6], + "date_range": [datetime(2020, 2, 24, 0, minute) for minute in range(4)], + "date_obj0": ["2020-02-24", "2020-02-25", "2020-02-26", "2020-02-27"], +} + +# a is uncorrelated, b is correlated with c and d, and c and d only with b. +DATA_CORR = { + "a": [1, -1, 0, 0, 0, 0, 2], + "b": [0, 0, 1, -1, 1, -1, 0], + "c": [0, 0, 1, -1, 0, 0, 1], + "d": [0, 0, 0, 0, 1, -1, 1], +} + +EXPECTED_MEAN = { + "var_0": 0.5813469607144305, + "var_1": 0.5325152703164752, + "var_2": 0.5023573007759755, + "var_3": 0.47596844810700234, + "var_4": 0.9696712897767115, + "var_5": 0.5078009005719849, + "var_6": 0.966096275433625, + "var_7": 0.9918595739378872, + "var_8": 0.521667767752105, + "var_9": 0.9476311088509884, + "var_10": 0.4871054926777818, + "var_11": 0.5180029642379039, +} + +EXPECTED_STD = { + "var_0": 0.0035430274728173775, + "var_1": 0.0046697767238672565, + "var_2": 0.023714708852568194, + "var_3": 0.04219857610624132, + "var_4": 0.010364079344188424, + "var_5": 0.03203946151605523, + "var_6": 0.0063709642968091335, + "var_7": 0.0014579159677989356, + "var_8": 0.027570153897628277, + "var_9": 0.014363240810578251, + "var_10": 0.020283618255582142, + "var_11": 0.02707242215734807, +} + + +EXPECTED_IMPORTANCE_MEAN = { + "var_0": 0.008110472647428566, + "var_1": 0.004425867029009318, + "var_2": 0.0014110527658847542, + "var_3": 0.0, + "var_4": 0.09519163119147249, + "var_5": 0.005151538162222261, + "var_6": 0.06819196935501609, + "var_7": 0.7958920351591532, + "var_8": 0.005514122712728161, + "var_9": 0.006699116878609683, + "var_10": 0.0006441114577744834, + "var_11": 0.008768082640701015, +} + +EXPECTED_IMPORTANCE_STD = { + "var_0": 0.0044896289532760465, + "var_1": 0.004400023300047043, + "var_2": 0.0012779435856766451, + "var_3": 0.0, + "var_4": 0.15831114958148926, + "var_5": 0.005860881861755391, + "var_6": 0.0666073858478959, + "var_7": 0.12339701768095801, + "var_8": 0.0037045342320885986, + "var_9": 0.0016309720229483885, + "var_10": 0.0011156337706026607, + "var_11": 0.005458498614789974, +} + + +def _split_target(make_df, data): + X = make_df({k: v for k, v in data.items() if k != "target"}) + return X, make_series(make_df, data["target"]) + + +def _pearson(x, y): + return np.corrcoef(x, y)[0, 1] + + +@pytest.fixture(scope="module") +def data_with_groups(): + rng = np.random.default_rng(1) + data = {f"var_{i}": rng.normal(size=100).tolist() for i in range(1, 6)} + data["target"] = rng.integers(0, 100, size=100).tolist() + groups = np.repeat(np.arange(1, 11), 10) + rng.shuffle(groups) + return data, groups.tolist() + + +@pytest.mark.parametrize( + "variables, confirm_variables, exclude_datetime, expected", + [ + (None, False, False, list(DATA_VARTYPES)), + (None, False, True, ["Name", "City", "Age", "Marks"]), + (["Name", "Age"], False, True, ["Name", "Age"]), + (["Name", "Age", "Hola"], True, True, ["Name", "Age"]), + ], +) +def test_select_all_variables( + make_df, variables, confirm_variables, exclude_datetime, expected +): + variables_ = _select_all_variables( + make_df(DATA_VARTYPES), variables, confirm_variables, exclude_datetime ) - df["Name"] = df["Name"].astype("category") - return df + assert variables_ == expected -def test_select_all_variables(df): - # select all variables - assert ( - _select_all_variables( - df, variables=None, confirm_variables=False, exclude_datetime=False - ) - == df.columns.to_list() +@pytest.mark.parametrize( + "variables, confirm_variables, expected", + [ + (None, False, ["Age", "Marks"]), + (["Marks"], False, ["Marks"]), + (["Marks", "Hola"], True, ["Marks"]), + ], +) +def test_select_numerical_variables(make_df, variables, confirm_variables, expected): + variables_ = _select_numerical_variables( + make_df(DATA_VARTYPES), variables, confirm_variables ) + assert variables_ == expected - # select all variables except datetime - assert _select_all_variables( - df, variables=None, confirm_variables=False, exclude_datetime=True - ) == ["Name", "City", "Age", "Marks"] - - # select subset of variables, without confirm - subset = ["Name", "City", "Age", "Marks"] - assert ( - _select_all_variables( - df, variables=subset, confirm_variables=False, exclude_datetime=True - ) - == subset - ) - # select subset of variables, with confirm - subset = ["Name", "City", "Age", "Marks", "Hola"] - assert ( - _select_all_variables( - df, variables=subset, confirm_variables=True, exclude_datetime=True - ) - == subset[:-1] +@pytest.mark.parametrize( + "variables, expected", + [ + (["a", "b", "c", "d"], ([{"b", "c", "d"}], ["c", "d"], {"b": {"c", "d"}})), + (["a", "c", "b", "d"], ([{"c", "b"}], ["b"], {"c": {"b"}})), + ], +) +def test_find_correlated_features(make_df, variables, expected): + X = make_df( + { + "a": [1, -1, 0, 0, 0, 0], + "b": [0, 0, 1, -1, 1, -1], + "c": [0, 0, 1, -1, 0, 0], + "d": [0, 0, 0, 0, 1, -1], + } ) + assert find_correlated_features(X, variables, "pearson", 0.7) == expected -def test_select_numerical_variables(df): - # select all numerical variables - assert _select_numerical_variables( - df, - variables=None, - confirm_variables=False, - ) == ["Age", "Marks"] - - # select subset of variables, without confirm - subset = ["Marks"] - assert ( - _select_numerical_variables( - df, - variables=subset, - confirm_variables=False, - ) - == subset - ) +@pytest.mark.parametrize("method", ["pearson", "spearman", "kendall", _pearson]) +def test_find_correlated_features_methods(make_df, method): + X = make_df(DATA_CORR) + groups, drop, dict_ = find_correlated_features(X, list(DATA_CORR), method, 0.5) + assert groups == [{"b", "c", "d"}] + assert drop == ["c", "d"] + assert dict_ == {"b": {"c", "d"}} - # select subset of variables, with confirm - subset = ["Marks", "Hola"] - assert ( - _select_numerical_variables( - df, - variables=subset, - confirm_variables=True, - ) - == subset[:-1] - ) +@pytest.mark.parametrize( + "method, expected", + [ + ("pearson", ([{"a", "c"}, {"b", "d"}], ["c", "d"], {"a": {"c"}, "b": {"d"}})), + ("spearman", ([{"b", "d"}], ["d"], {"b": {"d"}})), + ("kendall", ([{"b", "d"}], ["d"], {"b": {"d"}})), + (_pearson, ([{"a", "c"}, {"b", "d"}], ["c", "d"], {"a": {"c"}, "b": {"d"}})), + ], +) +def test_find_correlated_features_skips_missing_values(make_df, method, expected): + # each pair of variables is compared on the rows where neither is missing. + data = { + "a": [None, -1, 0, 0, 0, 0, 2], + "b": [0, 0, 1, -1, 1, -1, None], + "c": [0, 0, None, -1, 0, 0, 1], + "d": [0, 0, 0, 0, 1, -1, 1], + } + X = make_df(data) + assert find_correlated_features(X, list(data), method, 0.6) == expected -def test_find_correlated_features(): - # given a correlation-threshold of 0.7 - # a is uncorrelated, - # b is collinear to c and d, - # c and d are collinear only to b. - X = pd.DataFrame() - X["a"] = [1, -1, 0, 0, 0, 0] - X["b"] = [0, 0, 1, -1, 1, -1] - X["c"] = [0, 0, 1, -1, 0, 0] - X["d"] = [0, 0, 0, 0, 1, -1] +@pytest.mark.parametrize("method", ["pearson", "spearman", "kendall"]) +def test_find_correlated_features_ignores_constant_variables(make_df, method): + X = make_df({**DATA_CORR, "e": [1] * 7}) groups, drop, dict_ = find_correlated_features( - X, variables=["a", "b", "c", "d"], method="pearson", threshold=0.7 + X, ["e", *DATA_CORR], method, 0.5 ) - assert groups == [{"b", "c", "d"}] assert drop == ["c", "d"] assert dict_ == {"b": {"c", "d"}} - groups, drop, dict_ = find_correlated_features( - X, variables=["a", "c", "b", "d"], method="pearson", threshold=0.7 - ) - assert groups == [{"c", "b"}] - assert drop == ["b"] - assert dict_ == {"c": {"b"}} +def test_find_correlated_features_with_integer_column_names(): + X = pd.DataFrame({i: values for i, values in enumerate(DATA_CORR.values())}) + groups, drop, dict_ = find_correlated_features(X, [0, 1, 2, 3], "pearson", 0.5) + assert groups == [{1, 2, 3}] + assert drop == [2, 3] + assert dict_ == {1: {2, 3}} -def test_single_feature_performance(df_test): - X, y = df_test +def test_single_feature_performance(make_df, data_classification): + X, y = _split_target(make_df, data_classification) rf = RandomForestClassifier(n_estimators=5, random_state=1) - variables = X.columns.to_list() mean_, std_ = single_feature_performance( X=X, y=y, - variables=variables, + variables=list(EXPECTED_MEAN), estimator=rf, cv=3, scoring="roc_auc", ) - - expected_mean = { - "var_0": 0.5813469607144305, - "var_1": 0.5325152703164752, - "var_2": 0.5023573007759755, - "var_3": 0.47596844810700234, - "var_4": 0.9696712897767115, - "var_5": 0.5078009005719849, - "var_6": 0.966096275433625, - "var_7": 0.9918595739378872, - "var_8": 0.521667767752105, - "var_9": 0.9476311088509884, - "var_10": 0.4871054926777818, - "var_11": 0.5180029642379039, - } - expected_std = { - "var_0": 0.0035430274728173775, - "var_1": 0.0046697767238672565, - "var_2": 0.023714708852568194, - "var_3": 0.04219857610624132, - "var_4": 0.010364079344188424, - "var_5": 0.03203946151605523, - "var_6": 0.0063709642968091335, - "var_7": 0.0014579159677989356, - "var_8": 0.027570153897628277, - "var_9": 0.014363240810578251, - "var_10": 0.020283618255582142, - "var_11": 0.02707242215734807, - } - assert mean_ == expected_mean - assert std_ == expected_std + assert mean_ == pytest.approx(EXPECTED_MEAN) + assert std_ == pytest.approx(EXPECTED_STD) -def test_single_feature_performance_cv_generator(df_test): - X, y = df_test +@pytest.mark.parametrize("cv_type", ["splitter", "generator"]) +def test_single_feature_performance_with_cv_splitter( + make_df, data_classification, cv_type +): + X, y = _split_target(make_df, data_classification) rf = RandomForestClassifier(n_estimators=5, random_state=1) - variables = X.columns.to_list() cv = StratifiedKFold(n_splits=3) - for cv_ in [cv, cv.split(X, y)]: - mean_, _ = single_feature_performance( - X=X, - y=y, - variables=variables, - estimator=rf, - cv=cv_, - scoring="roc_auc", - ) - - expected_mean = { - "var_0": 0.5813469607144305, - "var_1": 0.5325152703164752, - "var_2": 0.5023573007759755, - "var_3": 0.47596844810700234, - "var_4": 0.9696712897767115, - "var_5": 0.5078009005719849, - "var_6": 0.966096275433625, - "var_7": 0.9918595739378872, - "var_8": 0.521667767752105, - "var_9": 0.9476311088509884, - "var_10": 0.4871054926777818, - "var_11": 0.5180029642379039, - } - assert mean_ == expected_mean + if cv_type == "generator": + cv = cv.split(X, y) + + mean_, _ = single_feature_performance( + X=X, y=y, variables=list(EXPECTED_MEAN), estimator=rf, cv=cv, scoring="roc_auc" + ) + assert mean_ == pytest.approx(EXPECTED_MEAN) -def test_single_feature_performance_with_groups(df_test_with_groups): - X, y, groups = df_test_with_groups +@pytest.mark.parametrize("target_type", [list, np.array]) +def test_single_feature_performance_with_list_and_array_target( + make_df, data_classification, target_type +): + X, _ = _split_target(make_df, data_classification) + y = target_type(data_classification["target"]) rf = RandomForestClassifier(n_estimators=5, random_state=1) - variables = X.columns.to_list() - scoring = "neg_mean_absolute_error" + + mean_, _ = single_feature_performance( + X=X, y=y, variables=["var_0", "var_4"], estimator=rf, cv=3, scoring="roc_auc" + ) + assert mean_ == pytest.approx( + {"var_0": EXPECTED_MEAN["var_0"], "var_4": EXPECTED_MEAN["var_4"]} + ) + + +def test_single_feature_performance_with_groups(make_df, data_with_groups): + data, groups = data_with_groups + X, y = _split_target(make_df, data) + rf = RandomForestClassifier(n_estimators=5, random_state=1) + variables = ["var_1", "var_2", "var_3", "var_4", "var_5"] cv = GroupKFold(n_splits=3) - cv_indices = cv.split(X=X, y=y, groups=groups) expected_mean_, expected_std_ = single_feature_performance( X=X, y=y, variables=variables, estimator=rf, - cv=cv_indices, - scoring=scoring, + cv=cv.split(X=X, y=y, groups=groups), + scoring="neg_mean_absolute_error", ) - mean_, std_ = single_feature_performance( X=X, y=y, variables=variables, estimator=rf, cv=cv, - scoring=scoring, + scoring="neg_mean_absolute_error", groups=groups, ) - assert mean_ == expected_mean_ assert std_ == expected_std_ -def test_find_feature_importance(df_test): - X, y = df_test +@pytest.mark.parametrize("cv_type", ["splitter", "generator"]) +def test_find_feature_importance(make_df, data_classification, cv_type): + X, y = _split_target(make_df, data_classification) rf = RandomForestClassifier(n_estimators=3, random_state=3) cv = StratifiedKFold(n_splits=3) - scoring = "recall" - - expected_mean = pd.Series( - data=[0.01, 0.0, 0.0, 0.0, 0.1, 0.01, 0.07, 0.8, 0.01, 0.01, 0.0, 0.01], - index=[ - "var_0", - "var_1", - "var_2", - "var_3", - "var_4", - "var_5", - "var_6", - "var_7", - "var_8", - "var_9", - "var_10", - "var_11", - ], - ) - expected_std = pd.Series( - data=[ - 0.0045, - 0.0044, - 0.0013, - 0.0, - 0.1583, - 0.0059, - 0.0666, - 0.1234, - 0.0037, - 0.0016, - 0.0011, - 0.0055, - ], - index=[ - "var_0", - "var_1", - "var_2", - "var_3", - "var_4", - "var_5", - "var_6", - "var_7", - "var_8", - "var_9", - "var_10", - "var_11", - ], - ) + if cv_type == "generator": + cv = cv.split(X, y) mean_, std_ = find_feature_importance( - X=X, - y=y, - estimator=rf, - cv=cv, - scoring=scoring, + X=X, y=y, estimator=rf, cv=cv, scoring="recall" ) - pd.testing.assert_series_equal(mean_.round(2), expected_mean) - pd.testing.assert_series_equal(std_.round(4), expected_std) - mean_, std_ = find_feature_importance( - X=X, - y=y, - estimator=rf, - cv=cv.split(X, y), - scoring=scoring, - ) - pd.testing.assert_series_equal(mean_.round(2), expected_mean) - pd.testing.assert_series_equal(std_.round(4), expected_std) + importance_type = pd.Series if make_df is pd.DataFrame else dict + assert isinstance(mean_, importance_type) + assert isinstance(std_, importance_type) + assert dict(mean_) == pytest.approx(EXPECTED_IMPORTANCE_MEAN) + assert dict(std_) == pytest.approx(EXPECTED_IMPORTANCE_STD) -def test_find_feature_importancewith_groups(df_test_with_groups): - X, y, groups = df_test_with_groups +def test_find_feature_importance_with_groups(make_df, data_with_groups): + data, groups = data_with_groups + X, y = _split_target(make_df, data) rf = RandomForestClassifier(n_estimators=3, random_state=1) cv = GroupKFold(n_splits=3) - scoring = "neg_mean_absolute_error" - cv_indices = cv.split(X=X, y=y, groups=groups) expected_mean_, expected_std_ = find_feature_importance( X=X, y=y, estimator=rf, - cv=cv_indices, - scoring=scoring, + cv=cv.split(X=X, y=y, groups=groups), + scoring="neg_mean_absolute_error", ) - mean_, std_ = find_feature_importance( X=X, y=y, estimator=rf, cv=cv, - scoring=scoring, - groups=groups + scoring="neg_mean_absolute_error", + groups=groups, ) + assert dict(mean_) == dict(expected_mean_) + assert dict(std_) == dict(expected_std_) - pd.testing.assert_series_equal(mean_, expected_mean_) - pd.testing.assert_series_equal(std_, expected_std_) + +def test_find_feature_importance_returns_series_indexed_by_columns(): + X = pd.DataFrame({"a": [1.0, 2, 3, 4, 5, 6], "b": [0.5, 0, 1, 1, 3, 2]}) + X.columns.name = "features" + y = pd.Series([1.0, 2, 3, 4, 5, 6]) + + mean_, std_ = find_feature_importance( + X=X, y=y, estimator=Lasso(alpha=0.01), cv=2, scoring="r2" + ) + assert list(mean_.index) == ["a", "b"] + assert mean_.index.name == "features" + assert std_.index.name == "features" + + +@pytest.mark.parametrize( + "estimator, expected", + [ + (LogisticRegression(), [0.728 ** (1 / 3), 1.0]), + (Lasso(), [0.5, 2.0]), + ], +) +def test_get_feature_importances(estimator, expected): + if isinstance(estimator, Lasso): + estimator.coef_ = np.array([-0.5, 2.0]) + else: + estimator.coef_ = np.array([[0.6, 0.0], [0.0, 1.0], [0.8, 0.0]]) + assert get_feature_importances(estimator) == pytest.approx(expected) diff --git a/tests/test_selection/test_base_selector.py b/tests/test_selection/test_base_selector.py index 54cc5bfc0..7ef31226b 100644 --- a/tests/test_selection/test_base_selector.py +++ b/tests/test_selection/test_base_selector.py @@ -1,55 +1,163 @@ +import re +from datetime import datetime + +import numpy as np +import pandas as pd import pytest -from pandas.testing import assert_frame_equal +from sklearn.exceptions import NotFittedError from feature_engine.selection.base_selector import BaseSelector +from tests.backend_helpers import frame_to_dict + +DATA = { + "Name": ["tom", "nick", "krish", None], + "City": ["London", "Manchester", "Liverpool", "Bristol"], + "Age": [20, 21, 19, 18], + "Marks": [0.9, None, 0.7, 0.6], + "dob": [datetime(2020, 2, 24, 0, minute) for minute in range(4)], +} + + +# init parameters +@pytest.mark.parametrize("confirm_variables", [None, "hola", [True], 1, 0.5]) +def test_error_if_confirm_variables_not_bool(confirm_variables): + msg = ( + "confirm_variables takes only values True and False. " + f"Got {confirm_variables} instead." + ) + with pytest.raises(ValueError, match=re.escape(msg)): + BaseSelector(confirm_variables=confirm_variables) -@pytest.mark.parametrize("val", [None, "hola", [True]]) -def test_confirm_variables_in_init(val): - with pytest.raises(ValueError): - BaseSelector(confirm_variables=val) +@pytest.mark.parametrize("confirm_variables", [True, False]) +def test_init_param_assignment(confirm_variables): + selector = BaseSelector(confirm_variables=confirm_variables) + assert selector.confirm_variables is confirm_variables -class MockClass(BaseSelector): - def __init__(self, variables=None, confirm_variables=False): - self.variables = variables - self.confirm_variables = confirm_variables +# fit and transform +class MockSelector(BaseSelector): + def __init__(self, features_to_drop=("Name", "Marks")): + self.features_to_drop = features_to_drop + self.confirm_variables = False def fit(self, X, y=None): - self.features_to_drop_ = ["Name", "Marks"] + self.features_to_drop_ = list(self.features_to_drop) self._get_feature_names_in(X) return self -def test_transform_method(df_vartypes): - transformer = MockClass() - transformer.fit(df_vartypes) - Xt = transformer.transform(df_vartypes) +def test_transform_drops_features(make_df): + X = make_df(DATA) + Xt = MockSelector().fit(X).transform(X) + + assert isinstance(Xt, make_df) + assert frame_to_dict(Xt) == { + "City": ["London", "Manchester", "Liverpool", "Bristol"], + "Age": [20, 21, 19, 18], + "dob": DATA["dob"], + } + + +def test_transform_restores_train_column_order(make_df): + selector = MockSelector().fit(make_df(DATA)) + X = make_df({var: DATA[var] for var in ["dob", "Marks", "Age", "Name", "City"]}) + Xt = selector.transform(X) + + assert isinstance(Xt, make_df) + assert list(Xt.columns) == ["City", "Age", "dob"] + + +@pytest.mark.parametrize( + "features_to_drop, expected", + [ + ([], ["Name", "City", "Age", "Marks", "dob"]), + (["dob"], ["Name", "City", "Age", "Marks"]), + (["Age", "Name", "City", "dob"], ["Marks"]), + ], +) +def test_transform_returns_retained_features(make_df, features_to_drop, expected): + X = make_df(DATA) + Xt = MockSelector(features_to_drop).fit(X).transform(X) + + assert isinstance(Xt, make_df) + assert list(Xt.columns) == expected + + +def test_transform_does_not_modify_input(make_df): + X = make_df(DATA) + MockSelector().fit(X).transform(X) + assert frame_to_dict(X) == frame_to_dict(make_df(DATA)) + - # tests output of transform - assert_frame_equal(Xt, df_vartypes.drop(["Name", "Marks"], axis=1)) +def test_error_if_transform_df_has_different_number_of_columns(make_df): + selector = MockSelector().fit(make_df(DATA)) + msg = ( + "The number of columns in this dataset is different from the one used to " + "fit this transformer (when using the fit() method)." + ) + with pytest.raises(ValueError, match=re.escape(msg)): + selector.transform(make_df({"Age": DATA["Age"], "Marks": DATA["Marks"]})) + + +def test_error_if_transform_before_fit(make_df): + msg = ( + "This MockSelector instance is not fitted yet. Call 'fit' with " + "appropriate arguments before using this estimator." + ) + with pytest.raises(NotFittedError, match=re.escape(msg)): + MockSelector().transform(make_df(DATA)) - # tests this line: X = X[self.feature_names_in_] - assert_frame_equal( - transformer.transform(df_vartypes[["City", "Age", "Name", "Marks", "dob"]]), - Xt, + +def test_error_if_transform_input_not_dataframe(make_df): + selector = MockSelector().fit(make_df(DATA)) + msg = ( + "X must be a dataframe from a library supported by narwhals " + "(e.g. pandas, polars, PyArrow). Got instead." ) - # test error when there is a df shape missmatch - with pytest.raises(ValueError): - assert transformer.transform(df_vartypes[["Age", "Marks"]]) + with pytest.raises(TypeError, match=re.escape(msg)): + selector.transform(np.ones((4, 5))) + + +def test_get_feature_names_in(make_df): + selector = MockSelector() + selector._get_feature_names_in(make_df(DATA)) + assert selector.feature_names_in_ == ["Name", "City", "Age", "Marks", "dob"] + assert selector.n_features_in_ == 5 + + +def test_get_support(make_df): + selector = MockSelector().fit(make_df(DATA)) + assert selector.get_support() == [False, True, True, False, True] + assert list(selector.get_support(indices=True)) == [1, 2, 4] + + +def test_get_feature_names_out(make_df): + selector = MockSelector().fit(make_df(DATA)) + assert selector.get_feature_names_out() == ["City", "Age", "dob"] + + +def test_check_variable_number(): + selector = MockSelector() + selector.variables_ = ["Age"] + msg = ( + "The selector needs at least 2 or more variables to select from. " + "Got only 1 variable: ['Age']." + ) + with pytest.raises(ValueError, match=re.escape(msg)): + selector._check_variable_number() + +def test_transform_with_integer_column_names(): + X = pd.DataFrame({i: values for i, values in enumerate(DATA.values())}) + selector = MockSelector(features_to_drop=[0, 3]).fit(X) + Xt = selector.transform(X[[4, 3, 2, 1, 0]]) -def test_get_feature_names_in(df_vartypes): - tr = MockClass() - tr._get_feature_names_in(df_vartypes) - assert tr.n_features_in_ == df_vartypes.shape[1] - assert tr.feature_names_in_ == list(df_vartypes.columns) + assert selector.feature_names_in_ == [0, 1, 2, 3, 4] + pd.testing.assert_frame_equal(Xt, X[[1, 2, 4]]) -def test_get_support(df_vartypes): - tr = MockClass() - tr.fit(df_vartypes) - v_bool = [False, True, True, False, True] - v_ind = [1, 2, 4] - assert tr.get_support() == v_bool - assert list(tr.get_support(indices=True)) == v_ind +def test_transform_keeps_pandas_index(): + X = pd.DataFrame(DATA, index=[10, 11, 12, 13]) + Xt = MockSelector().fit(X).transform(X) + pd.testing.assert_frame_equal(Xt, X[["City", "Age", "dob"]]) diff --git a/tests/test_selection/test_mrmr.py b/tests/test_selection/test_mrmr.py index d27d4ba8c..ac914a7fd 100644 --- a/tests/test_selection/test_mrmr.py +++ b/tests/test_selection/test_mrmr.py @@ -1,322 +1,453 @@ +import re + import numpy as np import pandas as pd import pytest - -from sklearn.datasets import ( - make_classification, - load_breast_cancer, - fetch_california_housing, -) -from sklearn.ensemble import RandomForestClassifier, RandomForestRegressor -from sklearn.feature_selection import ( - f_classif, - f_regression, - mutual_info_classif, - mutual_info_regression, -) -from sklearn.model_selection import GridSearchCV +from sklearn.exceptions import NotFittedError +from sklearn.model_selection import KFold, StratifiedKFold from feature_engine.selection import MRMR +from tests.backend_helpers import frame_to_dict, make_series + +VARIABLES = [f"var_{i}" for i in range(12)] +RF_GRID = {"max_depth": [2, 3], "n_estimators": [10]} + +MI_CLASSIF = [ + 0.052936601624909096, + 0.012431039630421026, + 0.0031179453717387062, + 0.0004679804673619614, + 0.5770540935952642, + 0.011045380844762365, + 0.5728270276712453, + 0.6566020567069197, + 0.012040401687987368, + 0.5035424468925758, + 0.0, + 0.0, +] +MI_REGRESSION = [ + 0.4827568966040907, + 0.1498525858592843, + 0.0, + 0.04212366814778079, + 0.1824019466447857, + 0.011297428359838158, + 0.19292720067399838, + 0.2236206517399788, + 0.5277855986492561, + 0.22394827371463455, + 0.0, + 0.002100737962146937, +] +F_CLASSIF = [ + 79.04849944065174, + 7.852911596746189, + 1.4628293436965305, + 0.7454302324350504, + 2360.5567490946137, + 0.78416950052947, + 2296.80729048166, + 5206.0776362490415, + 0.059549527394634295, + 1739.598743686081, + 0.0453654613398977, + 0.5613535054454387, +] +F_REGRESSION = [ + 1426.4396061315033, + 424.0297119316108, + 3.060543877897646, + 65.31838343563534, + 100.41062036724897, + 1.622842273663943, + 105.20415656422067, + 4.142473196617828, + 1683.2517259758326, + 160.1177438792669, + 0.07650822171810838, + 0.18645998061038163, +] +RF_CLASSIF = [ + 0.022247324184615325, + 0.003955855934811008, + 0.0011991561893078399, + 0.00026914064555155195, + 0.19395815835718982, + 0.0004983747672734107, + 0.09261141003009202, + 0.41634361330706404, + 0.004294463160439444, + 0.26440681633767793, + 0.00019819544535523348, + 1.7491640622450164e-05, +] +RF_REGRESSION = [ + 0.08500311779594995, + 0.21949616499172103, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.0, + 0.695500717212329, + 0.0, + 0.0, + 0.0, +] @pytest.fixture(scope="module") -def df_test_regression(): - X, y = make_classification( - n_samples=1000, - n_features=12, - n_redundant=4, - n_clusters_per_class=1, - weights=[0.50], - class_sep=2, - random_state=1, - ) - - # transform arrays into pandas df and series - colnames = ["var_" + str(i) for i in range(12)] - X = pd.DataFrame(X, columns=colnames) - y = pd.Series(0.5 * X["var_1"] - 0.3 * X["var_3"] + X["var_8"]) - return X, y +def data_regression(data_classification): + data = {k: v for k, v in data_classification.items() if k != "target"} + data["target"] = [ + 0.5 * a - 0.3 * b + c + for a, b, c in zip(data["var_1"], data["var_3"], data["var_8"]) + ] + return data -@pytest.mark.parametrize("method", ["MIQ", "MID", "FCQ", "FCD", "RFCQ"]) -def test_method_param(method): - tr = MRMR(method=method) - assert tr.method == method +def _split_target(make_df, data): + X = make_df({k: v for k, v in data.items() if k != "target"}) + return X, make_series(make_df, data["target"]) -@pytest.mark.parametrize("method", [10, "string", False, [0, 1]]) -def test_method_raises_error(method): +# init parameters +@pytest.mark.parametrize("method", [10, "string", False, None, ["MIQ"]]) +def test_error_if_method_not_allowed(method): msg = ( "method must be one of 'MIQ', 'MID', 'FCQ', 'FCD', 'RFCQ'. " f"Got {method} instead." ) - with pytest.raises(ValueError) as record: + with pytest.raises(ValueError, match=re.escape(msg)): MRMR(method=method) - assert str(record.value) == msg - - -@pytest.mark.parametrize("max_features", [1, 10, 20, None]) -def test_max_features_param(max_features): - tr = MRMR(max_features=max_features) - if max_features is not None: - assert tr.max_features == max_features - else: - assert tr.max_features is None -@pytest.mark.parametrize("max_features", ["string", -1, [0, 1]]) -def test_max_features_raises_error(max_features): +@pytest.mark.parametrize("max_features", ["string", -1, 0, 1.5, [0, 1]]) +def test_error_if_max_features_not_positive_integer(max_features): msg = ( "max_features must be an integer with the number of features to " f"select. Got {max_features} instead." ) - with pytest.raises(ValueError) as record: + with pytest.raises(ValueError, match=re.escape(msg)): MRMR(max_features=max_features) - assert str(record.value) == msg @pytest.mark.parametrize( - ("vars", "max_feat"), + "variables, max_features", [(["var1", "var2"], 2), (["var1", "var2"], 3), (["var1", "var2", "var3"], 4)], ) -def test_max_features_raises_error_when_more_than_vars(vars, max_feat): +def test_error_if_max_features_not_less_than_number_of_variables( + variables, max_features +): msg = ( - f"The number of variables to examine is {len(vars)}, which is " + f"The number of variables to examine is {len(variables)}, which is " "less than or equal to the number of features to select indicated " - f"in `max_features`, which is {max_feat}. Please check the " + f"in `max_features`, which is {max_features}. Please check the " "values entered in the parameters `variables` and `max_features`." ) - with pytest.raises(ValueError) as record: - MRMR(variables=vars, max_features=max_feat) - assert str(record.value) == msg + with pytest.raises(ValueError, match=re.escape(msg)): + MRMR(variables=variables, max_features=max_features) @pytest.mark.parametrize("scoring", ["roc_auc", "accuracy", "precision"]) -def test_raises_error_when_metric_not_suitable_for_regression(scoring): +def test_error_if_scoring_not_for_regression(scoring): msg = ( f"The metric {scoring} is not suitable for regression. Set the " "parameter regression to False or choose a different performance " "metric." ) - with pytest.raises(ValueError) as record: + with pytest.raises(ValueError, match=re.escape(msg)): MRMR(method="RFCQ", regression=True, scoring=scoring) - assert str(record.value) == msg - - -@pytest.mark.parametrize( - "method", - [ - "MIQ", - "MID", - "FCQ", - "FCD", - ], -) -@pytest.mark.parametrize("scoring", ["roc_auc", "accuracy", "precision"]) -def test_metric_does_not_raise_error_when_not_RF_regression(method, scoring): - # scoring is only used when method='RFCQ'. Otherwise it should be ignored. - tr = MRMR(method=method, regression=True, scoring=scoring) - assert tr.method == method - assert tr.scoring == scoring @pytest.mark.parametrize("scoring", ["mse", "mae", "r2"]) -def test_raises_error_when_metric_not_suitable_for_classif(scoring): +def test_error_if_scoring_not_for_classification(scoring): msg = ( - f"The metric {scoring} is not suitable for classification. Set the" + f"The metric {scoring} is not suitable for classification. Set the " "parameter regression to True or choose a different performance " "metric." ) - with pytest.raises(ValueError) as record: + with pytest.raises(ValueError, match=re.escape(msg)): MRMR(method="RFCQ", regression=False, scoring=scoring) - assert str(record.value) == msg @pytest.mark.parametrize( - "method", + "method, max_features, discrete_features, n_neighbors, scoring, cv, " + "param_grid, regression, confirm_variables, random_state, n_jobs", [ - "MIQ", - "MID", - "FCQ", - "FCD", + ("MIQ", None, "auto", 3, "roc_auc", 3, None, False, False, None, None), + ("MID", 5, [True, False], 5, "r2", 5, None, True, True, 0, 2), + ("FCQ", 1, False, 1, "accuracy", KFold(), None, False, False, 42, -1), + ("FCD", 10, True, 3, "mse", 3, None, True, False, None, None), + ("RFCQ", 3, "auto", 3, "r2", 2, RF_GRID, True, True, 1, None), + ("RFCQ", 3, "auto", 3, "accuracy", 2, None, False, False, 1, None), ], ) -@pytest.mark.parametrize("scoring", ["mse", "mae", "r2"]) -def test_metric_does_not_raise_error_when_not_RF_classif(method, scoring): - # scoring is only used when method='RFCQ'. Otherwise it should be ignored. - tr = MRMR(method=method, regression=False, scoring=scoring) - assert tr.method == method - assert tr.scoring == scoring - - -@pytest.mark.parametrize("method", ["MID", "MIQ"]) -def test_calculate_relevance_mi(df_test, df_test_regression, method): - X, y = df_test - expected_relevance = mutual_info_classif(X, y, random_state=42) - - sel = MRMR(method=method, regression=False, random_state=42) - relevance = sel._calculate_relevance(X, y) - - assert np.array_equal(expected_relevance, relevance) - - X, y = df_test_regression - expected_relevance = mutual_info_regression(X, y, random_state=42) - - sel = MRMR(method=method, regression=True, random_state=42) - relevance = sel._calculate_relevance(X, y) - - assert np.array_equal(expected_relevance, relevance) - +def test_init_param_assignment( + method, + max_features, + discrete_features, + n_neighbors, + scoring, + cv, + param_grid, + regression, + confirm_variables, + random_state, + n_jobs, +): + sel = MRMR( + method=method, + max_features=max_features, + discrete_features=discrete_features, + n_neighbors=n_neighbors, + scoring=scoring, + cv=cv, + param_grid=param_grid, + regression=regression, + confirm_variables=confirm_variables, + random_state=random_state, + n_jobs=n_jobs, + ) + assert sel.method == method + assert sel.max_features == max_features + assert sel.discrete_features == discrete_features + assert sel.n_neighbors == n_neighbors + assert sel.scoring == scoring + assert sel.cv is cv + assert sel.param_grid == param_grid + assert sel.regression is regression + assert sel.confirm_variables is confirm_variables + assert sel.random_state == random_state + assert sel.n_jobs == n_jobs + + +# fit and transform +@pytest.mark.parametrize( + "method, relevance, selected", + [ + ("MIQ", MI_CLASSIF, ["var_6", "var_7", "var_10", "var_11"]), + ("MID", MI_CLASSIF, ["var_1", "var_4", "var_5", "var_7"]), + ("FCQ", F_CLASSIF, ["var_4", "var_6", "var_7", "var_9"]), + ("FCD", F_CLASSIF, ["var_4", "var_6", "var_7", "var_9"]), + ("RFCQ", RF_CLASSIF, ["var_0", "var_4", "var_7", "var_9"]), + ], +) +def test_fit_transform_classification( + make_df, data_classification, method, relevance, selected +): + X, y = _split_target(make_df, data_classification) + sel = MRMR( + method=method, + max_features=4, + param_grid=RF_GRID, + regression=False, + random_state=0, + ) + Xt = sel.fit_transform(X, y) -@pytest.mark.parametrize("method", ["FCD", "FCQ"]) -def test_calculate_relevance_f(df_test, df_test_regression, method): - X, y = df_test - expected_relevance = f_classif(X, y)[0] + assert sel.variables_ == VARIABLES + assert sel.relevance_.tolist() == pytest.approx(relevance) + assert sel.features_to_drop_ == [f for f in VARIABLES if f not in selected] + assert sel.feature_names_in_ == VARIABLES + assert sel.n_features_in_ == 12 + assert isinstance(Xt, make_df) + assert list(frame_to_dict(Xt)) == selected + assert frame_to_dict(Xt) == {f: data_classification[f] for f in selected} - sel = MRMR(method=method, regression=False, random_state=42) - relevance = sel._calculate_relevance(X, y) - assert np.array_equal(expected_relevance, relevance) +@pytest.mark.parametrize( + "method, relevance, selected", + [ + ("MIQ", MI_REGRESSION, ["var_1", "var_2", "var_5", "var_8"]), + ("MID", MI_REGRESSION, ["var_1", "var_3", "var_7", "var_8"]), + ("FCQ", F_REGRESSION, ["var_0", "var_1", "var_3", "var_8"]), + ("FCD", F_REGRESSION, ["var_0", "var_1", "var_8", "var_9"]), + ("RFCQ", RF_REGRESSION, ["var_0", "var_1", "var_2", "var_8"]), + ], +) +def test_fit_transform_regression( + make_df, data_regression, method, relevance, selected +): + X, y = _split_target(make_df, data_regression) + sel = MRMR( + method=method, + max_features=4, + scoring="r2", + param_grid=RF_GRID, + regression=True, + random_state=0, + ) + Xt = sel.fit_transform(X, y) - X, y = df_test_regression - expected_relevance = f_regression(X, y)[0] + assert sel.relevance_.tolist() == pytest.approx(relevance) + assert sel.features_to_drop_ == [f for f in VARIABLES if f not in selected] + assert isinstance(Xt, make_df) + assert frame_to_dict(Xt) == {f: data_regression[f] for f in selected} - sel = MRMR(method=method, regression=True, random_state=42) - relevance = sel._calculate_relevance(X, y) - assert np.array_equal(expected_relevance, relevance) +def test_default_max_features_selects_20_percent(make_df, data_classification): + X, y = _split_target(make_df, data_classification) + sel = MRMR(method="FCQ").fit(X, y) + assert sel.features_to_drop_ == [ + f for f in VARIABLES if f not in ["var_4", "var_7"] + ] -def test_calculate_relevance_RF(df_test, df_test_regression): - param_grid = {"max_depth": [1, 2, 3, 4]} +def test_variables_and_discrete_features(make_df, data_classification): + data = dict(data_classification) + data["var_1"] = [round(2 * v) for v in data["var_1"]] + data["var_5"] = [round(v) for v in data["var_5"]] + X, y = _split_target(make_df, data) + sel = MRMR( + variables=["var_0", "var_1", "var_4", "var_5", "var_7"], + method="MID", + max_features=3, + discrete_features=[False, True, False, True, False], + random_state=0, + ) + Xt = sel.fit_transform(X, y) + + assert sel.variables_ == ["var_0", "var_1", "var_4", "var_5", "var_7"] + assert sel.relevance_.tolist() == pytest.approx( + [ + 0.052936601624909096, + 0.01026964337663472, + 0.5770540935952642, + 0.0017639968720600564, + 0.6566020567069197, + ] + ) + assert sel.features_to_drop_ == ["var_0", "var_4"] + assert isinstance(Xt, make_df) + assert list(frame_to_dict(Xt)) == [ + f for f in VARIABLES if f not in ["var_0", "var_4"] + ] - X, y = df_test - model = RandomForestClassifier(random_state=42) - model = GridSearchCV(model, cv=3, scoring="roc_auc", param_grid=param_grid) - model.fit(X, y) - expected_relevance = model.best_estimator_.feature_importances_ +def test_confirm_variables(make_df, data_classification): + X, y = _split_target(make_df, data_classification) sel = MRMR( - method="RFCQ", regression=False, random_state=42, cv=3, scoring="roc_auc" + variables=["var_0", "var_2", "var_4", "var_6", "var_8", "var_20"], + method="MIQ", + max_features=3, + confirm_variables=True, + random_state=0, ) - relevance = sel._calculate_relevance(X, y) + sel.fit(X, y) - assert np.array_equal(expected_relevance, relevance) + assert sel.variables_ == ["var_0", "var_2", "var_4", "var_6", "var_8"] + assert sel.relevance_.tolist() == pytest.approx( + [ + 0.052936601624909096, + 0.0031179453717387062, + 0.5770540935952642, + 0.5728270276712453, + 0.012040401687987368, + ] + ) + assert sel.features_to_drop_ == ["var_0", "var_8"] - X, y = df_test_regression - model = RandomForestRegressor(random_state=42) - model = GridSearchCV(model, cv=3, scoring="r2", param_grid=param_grid) - model.fit(X, y) - expected_relevance = model.best_estimator_.feature_importances_ +def test_cv_generator(make_df, data_classification): + X, y = _split_target(make_df, data_classification) sel = MRMR( method="RFCQ", - regression=True, - random_state=42, - cv=3, - scoring="r2", - param_grid=param_grid, + max_features=3, + param_grid=RF_GRID, + cv=StratifiedKFold(n_splits=3).split(X, y), + random_state=0, ) - relevance = sel._calculate_relevance(X, y) - - assert np.array_equal(expected_relevance, relevance) - + sel.fit(X, y) -@pytest.mark.parametrize( - ("method1", "method2"), [("MID", "FCD"), ("MIQ", "FCQ"), ("MIQ", "RFCQ")] -) -def test_calculate_redundance(df_test_regression, method1, method2): - X, y = df_test_regression - - expected_redundance = mutual_info_regression(X, y, random_state=42) - sel = MRMR(method=method1, random_state=42) - redundance = sel._calculate_redundance(X, y) - assert np.array_equal(expected_redundance, redundance) - - expected_redundance = np.absolute(X.corrwith(y).values) - sel = MRMR(method=method2) - redundance = sel._calculate_redundance(X, y) - assert np.array_equal(expected_redundance, redundance) - - -@pytest.mark.parametrize("method", ["MID", "MIQ", "FCD", "FCQ", "RFCQ"]) -def test_calculate_mrmr(method): - relevance = np.array([1, 2, 3, 4, 5]) - redundance = np.array([1, 2, 3, 4, 5]) - - sel = MRMR(method=method) - mrmr = sel._calculate_mrmr(relevance, redundance) - - if method in ["MID", "FCD"]: - expected_mrmr = relevance - redundance - else: - expected_mrmr = relevance / redundance - assert np.array_equal(expected_mrmr, mrmr) - - -def test_mrmr_mid_and_miq_classif(): - X, y = load_breast_cancer(return_X_y=True, as_frame=True) - - selected = [ - "worst perimeter", - "fractal dimension error", - "worst texture", - "worst smoothness", - "mean concave points", - "perimeter error", - "worst symmetry", - "worst concave points", - "area error", - "mean perimeter", + assert sel.relevance_.tolist() == pytest.approx(RF_CLASSIF) + assert sel.features_to_drop_ == [ + f for f in VARIABLES if f not in ["var_4", "var_7", "var_9"] ] - to_drop = [f for f in X.columns if f not in selected] - sel = MRMR(method="MIQ", max_features=10, regression=False, random_state=42) - sel.fit(X, y) - Xtr = sel.transform(X) - - assert sel.features_to_drop_ == to_drop - pd.testing.assert_frame_equal(X.drop(to_drop, axis=1), Xtr) - - selected = [ - "worst perimeter", - "worst smoothness", - "worst texture", - "mean concave points", - "perimeter error", - "worst concavity", - "worst symmetry", - "area error", - "symmetry error", - "worst concave points", - ] - to_drop = [f for f in X.columns if f not in selected] - sel = MRMR(method="MID", max_features=10, regression=False, random_state=42) +def test_missing_values_with_random_forests(make_df, data_classification): + # only random forests take missing values; the redundance then uses the rows + # where both features have values. + data = dict(data_classification) + for var, step in [("var_2", 7), ("var_5", 5), ("var_9", 3)]: + data[var] = [None if i % step == 0 else v for i, v in enumerate(data[var])] + X, y = _split_target(make_df, data) + sel = MRMR(method="RFCQ", max_features=4, param_grid=RF_GRID, random_state=0) sel.fit(X, y) - Xtr = sel.transform(X) - - assert sel.features_to_drop_ == to_drop - pd.testing.assert_frame_equal(X.drop(to_drop, axis=1), Xtr) + assert sel.relevance_.tolist() == pytest.approx( + [ + 0.0076636392509130055, + 0.005491342491500148, + 0.0037204783072902247, + 0.0, + 0.2392943866283686, + 3.639947920429752e-05, + 0.17849998892103483, + 0.46742647746615623, + 0.0011218247849137536, + 0.09508246490768371, + 0.0016629977629352137, + 0.0, + ] + ) + assert sel.features_to_drop_ == [ + f for f in VARIABLES if f not in ["var_4", "var_6", "var_7", "var_10"] + ] -def test_mrmr_fcd_and_fcq_regression(): - X, y = fetch_california_housing(return_X_y=True, as_frame=True) - selected = ["MedInc", "Latitude", "HouseAge", "AveRooms", "AveOccup"] - to_drop = [f for f in X.columns if f not in selected] +@pytest.mark.parametrize("target_type", [list, np.array]) +def test_target_as_list_or_array(make_df, data_classification, target_type): + X, _ = _split_target(make_df, data_classification) + y = target_type(data_classification["target"]) + sel = MRMR(method="FCD", max_features=3).fit(X, y) - sel = MRMR(method="FCQ", max_features=5, regression=True) - Xtr = sel.fit_transform(X, y) + assert sel.relevance_.tolist() == pytest.approx(F_CLASSIF) + assert sel.features_to_drop_ == [ + f for f in VARIABLES if f not in ["var_4", "var_6", "var_7"] + ] - assert sel.features_to_drop_ == to_drop - pd.testing.assert_frame_equal(X.drop(to_drop, axis=1), Xtr) - selected = ["MedInc", "AveRooms", "Latitude", "HouseAge"] - to_drop = [f for f in X.columns if f not in selected] +def test_error_if_less_than_2_variables(make_df, data_classification): + X, y = _split_target(make_df, data_classification) + msg = ( + "The selector needs at least 2 or more variables to select from. " + "Got only 1 variable: ['var_0']." + ) + with pytest.raises(ValueError, match=re.escape(msg)): + MRMR(variables=["var_0"]).fit(X, y) - sel = MRMR(method="FCD", max_features=4, regression=True) - Xtr = sel.fit_transform(X, y) - assert sel.features_to_drop_ == to_drop - pd.testing.assert_frame_equal(X.drop(to_drop, axis=1), Xtr) +def test_error_if_transform_before_fit(make_df, data_classification): + X, _ = _split_target(make_df, data_classification) + msg = ( + "This MRMR instance is not fitted yet. Call 'fit' with appropriate " + "arguments before using this estimator." + ) + with pytest.raises(NotFittedError, match=re.escape(msg)): + MRMR().transform(X) + + +@pytest.mark.parametrize("method", ["MIQ", "FCQ"]) +def test_integer_column_names(data_classification, method): + data = {i: data_classification[f"var_{i}"] for i in range(12)} + X = pd.DataFrame(data) + y = pd.Series(data_classification["target"]) + sel = MRMR(method=method, max_features=4, random_state=0) + Xt = sel.fit_transform(X, y) + + expected = {"MIQ": [6, 7, 10, 11], "FCQ": [4, 6, 7, 9]}[method] + assert sel.variables_ == list(range(12)) + assert sel.features_to_drop_ == [i for i in range(12) if i not in expected] + pd.testing.assert_frame_equal(Xt, X[expected]) + + +def test_keeps_pandas_index(data_classification): + X, y = _split_target(pd.DataFrame, data_classification) + X.index = X.index + 100 + y.index = X.index + Xt = MRMR(method="FCQ", max_features=4).fit_transform(X, y) + pd.testing.assert_frame_equal(Xt, X[["var_4", "var_6", "var_7", "var_9"]])