Source code for featuretools.selection.selection

import pandas as pd
from woodwork.logical_types import Boolean, BooleanNullable


[docs]def remove_low_information_features(feature_matrix, features=None): """Select features that have at least 2 unique values and that are not all null Args: feature_matrix (:class:`pd.DataFrame`): DataFrame whose columns are feature names and rows are instances features (list[:class:`featuretools.FeatureBase`] or list[str], optional): List of features to select Returns: (feature_matrix, features) """ keep = [ c for c in feature_matrix if ( feature_matrix[c].nunique(dropna=False) > 1 and feature_matrix[c].dropna().shape[0] > 0 ) ] feature_matrix = feature_matrix[keep] if features is not None: features = [f for f in features if f.get_name() in feature_matrix.columns] return feature_matrix, features return feature_matrix
[docs]def remove_highly_null_features(feature_matrix, features=None, pct_null_threshold=0.95): """ Removes columns from a feature matrix that have higher than a set threshold of null values. Args: feature_matrix (:class:`pd.DataFrame`): DataFrame whose columns are feature names and rows are instances. features (list[:class:`featuretools.FeatureBase`] or list[str], optional): List of features to select. pct_null_threshold (float): If the percentage of NaN values in an input feature exceeds this amount, that feature will be considered highly-null. Defaults to 0.95. Returns: pd.DataFrame, list[:class:`.FeatureBase`]: The feature matrix and the list of generated feature definitions. Matches dfs output. If no feature list is provided as input, the feature list will not be returned. """ if pct_null_threshold < 0 or pct_null_threshold > 1: raise ValueError( "pct_null_threshold must be a float between 0 and 1, inclusive.", ) percent_null_by_col = (feature_matrix.isnull().mean()).to_dict() if pct_null_threshold == 0.0: keep = [ f_name for f_name, pct_null in percent_null_by_col.items() if pct_null <= pct_null_threshold ] else: keep = [ f_name for f_name, pct_null in percent_null_by_col.items() if pct_null < pct_null_threshold ] return _apply_feature_selection(keep, feature_matrix, features)
[docs]def remove_single_value_features( feature_matrix, features=None, count_nan_as_value=False, ): """Removes columns in feature matrix where all the values are the same. Args: feature_matrix (:class:`pd.DataFrame`): DataFrame whose columns are feature names and rows are instances. features (list[:class:`featuretools.FeatureBase`] or list[str], optional): List of features to select. count_nan_as_value (bool): If True, missing values will be counted as their own unique value. If set to False, a feature that has one unique value and all other data missing will be removed from the feature matrix. Defaults to False. Returns: pd.DataFrame, list[:class:`.FeatureBase`]: The feature matrix and the list of generated feature definitions. Matches dfs output. If no feature list is provided as input, the feature list will not be returned. """ unique_counts_by_col = feature_matrix.nunique( dropna=not count_nan_as_value, ).to_dict() keep = [ f_name for f_name, unique_count in unique_counts_by_col.items() if unique_count > 1 ] return _apply_feature_selection(keep, feature_matrix, features)
[docs]def remove_highly_correlated_features( feature_matrix, features=None, pct_corr_threshold=0.95, features_to_check=None, features_to_keep=None, ): """Removes columns in feature matrix that are highly correlated with another column. Note: We make the assumption that, for a pair of features, the feature that is further right in the feature matrix produced by ``dfs`` is the more complex one. The assumption does not hold if the order of columns in the feature matrix has changed from what ``dfs`` produces. Args: feature_matrix (:class:`pd.DataFrame`): DataFrame whose columns are feature names and rows are instances. If Woodwork is not initalized, will perform Woodwork initialization, which may result in slightly different types than those in the original feature matrix created by Featuretools. features (list[:class:`featuretools.FeatureBase`] or list[str], optional): List of features to select. pct_corr_threshold (float): The correlation threshold to be considered highly correlated. Defaults to 0.95. features_to_check (list[str], optional): List of column names to check whether any pairs are highly correlated. Will not check any other columns, meaning the only columns that can be removed are in this list. If null, defaults to checking all columns. features_to_keep (list[str], optional): List of colum names to keep even if correlated to another column. If null, all columns will be candidates for removal. Returns: pd.DataFrame, list[:class:`.FeatureBase`]: The feature matrix and the list of generated feature definitions. Matches dfs output. If no feature list is provided as input, the feature list will not be returned. For consistent results, do not change the order of features outputted by dfs. """ if feature_matrix.ww.schema is None: feature_matrix.ww.init() if pct_corr_threshold < 0 or pct_corr_threshold > 1: raise ValueError( "pct_corr_threshold must be a float between 0 and 1, inclusive.", ) if features_to_check is None: features_to_check = list(feature_matrix.columns) else: for f_name in features_to_check: assert ( f_name in feature_matrix.columns ), "feature named {} is not in feature matrix".format(f_name) if features_to_keep is None: features_to_keep = [] to_select = ["numeric", Boolean, BooleanNullable] fm = feature_matrix.ww[features_to_check] fm_to_check = fm.ww.select(include=to_select) dropped = set() columns_to_check = fm_to_check.columns # When two features are found to be highly correlated, # we drop the more complex feature # Columns produced later in dfs are more complex for i in range(len(columns_to_check) - 1, 0, -1): more_complex_name = columns_to_check[i] more_complex_col = fm_to_check[more_complex_name] # Convert boolean or Int64 column to be float64 if pd.api.types.is_bool_dtype(more_complex_col) or isinstance( more_complex_col.dtype, pd.Int64Dtype, ): more_complex_col = more_complex_col.astype("float64") for j in range(i - 1, -1, -1): less_complex_name = columns_to_check[j] less_complex_col = fm_to_check[less_complex_name] # Convert boolean or Int64 column to be float64 if pd.api.types.is_bool_dtype(less_complex_col) or isinstance( less_complex_col.dtype, pd.Int64Dtype, ): less_complex_col = less_complex_col.astype("float64") if abs(more_complex_col.corr(less_complex_col)) >= pct_corr_threshold: dropped.add(more_complex_name) break keep = [ f_name for f_name in feature_matrix.columns if (f_name in features_to_keep or f_name not in dropped) ] return _apply_feature_selection(keep, feature_matrix, features)
def _apply_feature_selection(keep, feature_matrix, features=None): new_matrix = feature_matrix[keep] new_feature_names = set(new_matrix.columns) if features is not None: new_features = [] for f in features: if f.number_output_features > 1: slices = [ f[i] for i in range(f.number_output_features) if f[i].get_name() in new_feature_names ] if len(slices) == f.number_output_features: new_features.append(f) else: new_features.extend(slices) else: if f.get_name() in new_feature_names: new_features.append(f) return new_matrix, new_features return new_matrix