from time import time from collections import namedtuple import warnings from scipy import stats import numpy as np from ..base import clone from ..exceptions import ConvergenceWarning from ..preprocessing import normalize from ..utils import (check_array, check_random_state, _safe_indexing, is_scalar_nan) from ..utils.validation import FLOAT_DTYPES, check_is_fitted from ..utils._mask import _get_mask from ._base import _BaseImputer from ._base import SimpleImputer from ._base import _check_inputs_dtype _ImputerTriplet = namedtuple('_ImputerTriplet', ['feat_idx', 'neighbor_feat_idx', 'estimator']) class IterativeImputer(_BaseImputer): """Multivariate imputer that estimates each feature from all the others. A strategy for imputing missing values by modeling each feature with missing values as a function of other features in a round-robin fashion. Read more in the :ref:`User Guide `. .. versionadded:: 0.21 .. note:: This estimator is still **experimental** for now: the predictions and the API might change without any deprecation cycle. To use it, you need to explicitly import ``enable_iterative_imputer``:: >>> # explicitly require this experimental feature >>> from sklearn.experimental import enable_iterative_imputer # noqa >>> # now you can import normally from sklearn.impute >>> from sklearn.impute import IterativeImputer Parameters ---------- estimator : estimator object, default=BayesianRidge() The estimator to use at each step of the round-robin imputation. If ``sample_posterior`` is True, the estimator must support ``return_std`` in its ``predict`` method. missing_values : int, np.nan, default=np.nan The placeholder for the missing values. All occurrences of `missing_values` will be imputed. For pandas' dataframes with nullable integer dtypes with missing values, `missing_values` should be set to `np.nan`, since `pd.NA` will be converted to `np.nan`. sample_posterior : boolean, default=False Whether to sample from the (Gaussian) predictive posterior of the fitted estimator for each imputation. Estimator must support ``return_std`` in its ``predict`` method if set to ``True``. Set to ``True`` if using ``IterativeImputer`` for multiple imputations. max_iter : int, default=10 Maximum number of imputation rounds to perform before returning the imputations computed during the final round. A round is a single imputation of each feature with missing values. The stopping criterion is met once `max(abs(X_t - X_{t-1}))/max(abs(X[known_vals])) < tol`, where `X_t` is `X` at iteration `t`. Note that early stopping is only applied if ``sample_posterior=False``. tol : float, default=1e-3 Tolerance of the stopping condition. n_nearest_features : int, default=None Number of other features to use to estimate the missing values of each feature column. Nearness between features is measured using the absolute correlation coefficient between each feature pair (after initial imputation). To ensure coverage of features throughout the imputation process, the neighbor features are not necessarily nearest, but are drawn with probability proportional to correlation for each imputed target feature. Can provide significant speed-up when the number of features is huge. If ``None``, all features will be used. initial_strategy : str, default='mean' Which strategy to use to initialize the missing values. Same as the ``strategy`` parameter in :class:`~sklearn.impute.SimpleImputer` Valid values: {"mean", "median", "most_frequent", or "constant"}. imputation_order : str, default='ascending' The order in which the features will be imputed. Possible values: "ascending" From features with fewest missing values to most. "descending" From features with most missing values to fewest. "roman" Left to right. "arabic" Right to left. "random" A random order for each round. skip_complete : boolean, default=False If ``True`` then features with missing values during ``transform`` which did not have any missing values during ``fit`` will be imputed with the initial imputation method only. Set to ``True`` if you have many features with no missing values at both ``fit`` and ``transform`` time to save compute. min_value : float or array-like of shape (n_features,), default=-np.inf Minimum possible imputed value. Broadcast to shape (n_features,) if scalar. If array-like, expects shape (n_features,), one min value for each feature. The default is `-np.inf`. .. versionchanged:: 0.23 Added support for array-like. max_value : float or array-like of shape (n_features,), default=np.inf Maximum possible imputed value. Broadcast to shape (n_features,) if scalar. If array-like, expects shape (n_features,), one max value for each feature. The default is `np.inf`. .. versionchanged:: 0.23 Added support for array-like. verbose : int, default=0 Verbosity flag, controls the debug messages that are issued as functions are evaluated. The higher, the more verbose. Can be 0, 1, or 2. random_state : int, RandomState instance or None, default=None The seed of the pseudo random number generator to use. Randomizes selection of estimator features if n_nearest_features is not None, the ``imputation_order`` if ``random``, and the sampling from posterior if ``sample_posterior`` is True. Use an integer for determinism. See :term:`the Glossary `. add_indicator : boolean, default=False If True, a :class:`MissingIndicator` transform will stack onto output of the imputer's transform. This allows a predictive estimator to account for missingness despite imputation. If a feature has no missing values at fit/train time, the feature won't appear on the missing indicator even if there are missing values at transform/test time. Attributes ---------- initial_imputer_ : object of type :class:`~sklearn.impute.SimpleImputer` Imputer used to initialize the missing values. imputation_sequence_ : list of tuples Each tuple has ``(feat_idx, neighbor_feat_idx, estimator)``, where ``feat_idx`` is the current feature to be imputed, ``neighbor_feat_idx`` is the array of other features used to impute the current feature, and ``estimator`` is the trained estimator used for the imputation. Length is ``self.n_features_with_missing_ * self.n_iter_``. n_iter_ : int Number of iteration rounds that occurred. Will be less than ``self.max_iter`` if early stopping criterion was reached. n_features_with_missing_ : int Number of features with missing values. indicator_ : :class:`~sklearn.impute.MissingIndicator` Indicator used to add binary indicators for missing values. ``None`` if add_indicator is False. random_state_ : RandomState instance RandomState instance that is generated either from a seed, the random number generator or by `np.random`. See Also -------- SimpleImputer : Univariate imputation of missing values. Examples -------- >>> import numpy as np >>> from sklearn.experimental import enable_iterative_imputer >>> from sklearn.impute import IterativeImputer >>> imp_mean = IterativeImputer(random_state=0) >>> imp_mean.fit([[7, 2, 3], [4, np.nan, 6], [10, 5, 9]]) IterativeImputer(random_state=0) >>> X = [[np.nan, 2, 3], [4, np.nan, 6], [10, np.nan, 9]] >>> imp_mean.transform(X) array([[ 6.9584..., 2. , 3. ], [ 4. , 2.6000..., 6. ], [10. , 4.9999..., 9. ]]) Notes ----- To support imputation in inductive mode we store each feature's estimator during the ``fit`` phase, and predict without refitting (in order) during the ``transform`` phase. Features which contain all missing values at ``fit`` are discarded upon ``transform``. References ---------- .. [1] `Stef van Buuren, Karin Groothuis-Oudshoorn (2011). "mice: Multivariate Imputation by Chained Equations in R". Journal of Statistical Software 45: 1-67. `_ .. [2] `S. F. Buck, (1960). "A Method of Estimation of Missing Values in Multivariate Data Suitable for use with an Electronic Computer". Journal of the Royal Statistical Society 22(2): 302-306. `_ """ def __init__(self, estimator=None, *, missing_values=np.nan, sample_posterior=False, max_iter=10, tol=1e-3, n_nearest_features=None, initial_strategy="mean", imputation_order='ascending', skip_complete=False, min_value=-np.inf, max_value=np.inf, verbose=0, random_state=None, add_indicator=False): super().__init__( missing_values=missing_values, add_indicator=add_indicator ) self.estimator = estimator self.sample_posterior = sample_posterior self.max_iter = max_iter self.tol = tol self.n_nearest_features = n_nearest_features self.initial_strategy = initial_strategy self.imputation_order = imputation_order self.skip_complete = skip_complete self.min_value = min_value self.max_value = max_value self.verbose = verbose self.random_state = random_state def _impute_one_feature(self, X_filled, mask_missing_values, feat_idx, neighbor_feat_idx, estimator=None, fit_mode=True): """Impute a single feature from the others provided. This function predicts the missing values of one of the features using the current estimates of all the other features. The ``estimator`` must support ``return_std=True`` in its ``predict`` method for this function to work. Parameters ---------- X_filled : ndarray Input data with the most recent imputations. mask_missing_values : ndarray Input data's missing indicator matrix. feat_idx : int Index of the feature currently being imputed. neighbor_feat_idx : ndarray Indices of the features to be used in imputing ``feat_idx``. estimator : object The estimator to use at this step of the round-robin imputation. If ``sample_posterior`` is True, the estimator must support ``return_std`` in its ``predict`` method. If None, it will be cloned from self._estimator. fit_mode : boolean, default=True Whether to fit and predict with the estimator or just predict. Returns ------- X_filled : ndarray Input data with ``X_filled[missing_row_mask, feat_idx]`` updated. estimator : estimator with sklearn API The fitted estimator used to impute ``X_filled[missing_row_mask, feat_idx]``. """ if estimator is None and fit_mode is False: raise ValueError("If fit_mode is False, then an already-fitted " "estimator should be passed in.") if estimator is None: estimator = clone(self._estimator) missing_row_mask = mask_missing_values[:, feat_idx] if fit_mode: X_train = _safe_indexing(X_filled[:, neighbor_feat_idx], ~missing_row_mask) y_train = _safe_indexing(X_filled[:, feat_idx], ~missing_row_mask) estimator.fit(X_train, y_train) # if no missing values, don't predict if np.sum(missing_row_mask) == 0: return X_filled, estimator # get posterior samples if there is at least one missing value X_test = _safe_indexing(X_filled[:, neighbor_feat_idx], missing_row_mask) if self.sample_posterior: mus, sigmas = estimator.predict(X_test, return_std=True) imputed_values = np.zeros(mus.shape, dtype=X_filled.dtype) # two types of problems: (1) non-positive sigmas # (2) mus outside legal range of min_value and max_value # (results in inf sample) positive_sigmas = sigmas > 0 imputed_values[~positive_sigmas] = mus[~positive_sigmas] mus_too_low = mus < self._min_value[feat_idx] imputed_values[mus_too_low] = self._min_value[feat_idx] mus_too_high = mus > self._max_value[feat_idx] imputed_values[mus_too_high] = self._max_value[feat_idx] # the rest can be sampled without statistical issues inrange_mask = positive_sigmas & ~mus_too_low & ~mus_too_high mus = mus[inrange_mask] sigmas = sigmas[inrange_mask] a = (self._min_value[feat_idx] - mus) / sigmas b = (self._max_value[feat_idx] - mus) / sigmas truncated_normal = stats.truncnorm(a=a, b=b, loc=mus, scale=sigmas) imputed_values[inrange_mask] = truncated_normal.rvs( random_state=self.random_state_) else: imputed_values = estimator.predict(X_test) imputed_values = np.clip(imputed_values, self._min_value[feat_idx], self._max_value[feat_idx]) # update the feature X_filled[missing_row_mask, feat_idx] = imputed_values return X_filled, estimator def _get_neighbor_feat_idx(self, n_features, feat_idx, abs_corr_mat): """Get a list of other features to predict ``feat_idx``. If self.n_nearest_features is less than or equal to the total number of features, then use a probability proportional to the absolute correlation between ``feat_idx`` and each other feature to randomly choose a subsample of the other features (without replacement). Parameters ---------- n_features : int Number of features in ``X``. feat_idx : int Index of the feature currently being imputed. abs_corr_mat : ndarray, shape (n_features, n_features) Absolute correlation matrix of ``X``. The diagonal has been zeroed out and each feature has been normalized to sum to 1. Can be None. Returns ------- neighbor_feat_idx : array-like The features to use to impute ``feat_idx``. """ if (self.n_nearest_features is not None and self.n_nearest_features < n_features): p = abs_corr_mat[:, feat_idx] neighbor_feat_idx = self.random_state_.choice( np.arange(n_features), self.n_nearest_features, replace=False, p=p) else: inds_left = np.arange(feat_idx) inds_right = np.arange(feat_idx + 1, n_features) neighbor_feat_idx = np.concatenate((inds_left, inds_right)) return neighbor_feat_idx def _get_ordered_idx(self, mask_missing_values): """Decide in what order we will update the features. As a homage to the MICE R package, we will have 4 main options of how to order the updates, and use a random order if anything else is specified. Also, this function skips features which have no missing values. Parameters ---------- mask_missing_values : array-like, shape (n_samples, n_features) Input data's missing indicator matrix, where "n_samples" is the number of samples and "n_features" is the number of features. Returns ------- ordered_idx : ndarray, shape (n_features,) The order in which to impute the features. """ frac_of_missing_values = mask_missing_values.mean(axis=0) if self.skip_complete: missing_values_idx = np.flatnonzero(frac_of_missing_values) else: missing_values_idx = np.arange(np.shape(frac_of_missing_values)[0]) if self.imputation_order == 'roman': ordered_idx = missing_values_idx elif self.imputation_order == 'arabic': ordered_idx = missing_values_idx[::-1] elif self.imputation_order == 'ascending': n = len(frac_of_missing_values) - len(missing_values_idx) ordered_idx = np.argsort(frac_of_missing_values, kind='mergesort')[n:] elif self.imputation_order == 'descending': n = len(frac_of_missing_values) - len(missing_values_idx) ordered_idx = np.argsort(frac_of_missing_values, kind='mergesort')[n:][::-1] elif self.imputation_order == 'random': ordered_idx = missing_values_idx self.random_state_.shuffle(ordered_idx) else: raise ValueError("Got an invalid imputation order: '{0}'. It must " "be one of the following: 'roman', 'arabic', " "'ascending', 'descending', or " "'random'.".format(self.imputation_order)) return ordered_idx def _get_abs_corr_mat(self, X_filled, tolerance=1e-6): """Get absolute correlation matrix between features. Parameters ---------- X_filled : ndarray, shape (n_samples, n_features) Input data with the most recent imputations. tolerance : float, default=1e-6 ``abs_corr_mat`` can have nans, which will be replaced with ``tolerance``. Returns ------- abs_corr_mat : ndarray, shape (n_features, n_features) Absolute correlation matrix of ``X`` at the beginning of the current round. The diagonal has been zeroed out and each feature's absolute correlations with all others have been normalized to sum to 1. """ n_features = X_filled.shape[1] if (self.n_nearest_features is None or self.n_nearest_features >= n_features): return None with np.errstate(invalid='ignore'): # if a feature in the neighboorhood has only a single value # (e.g., categorical feature), the std. dev. will be null and # np.corrcoef will raise a warning due to a division by zero abs_corr_mat = np.abs(np.corrcoef(X_filled.T)) # np.corrcoef is not defined for features with zero std abs_corr_mat[np.isnan(abs_corr_mat)] = tolerance # ensures exploration, i.e. at least some probability of sampling np.clip(abs_corr_mat, tolerance, None, out=abs_corr_mat) # features are not their own neighbors np.fill_diagonal(abs_corr_mat, 0) # needs to sum to 1 for np.random.choice sampling abs_corr_mat = normalize(abs_corr_mat, norm='l1', axis=0, copy=False) return abs_corr_mat def _initial_imputation(self, X, in_fit=False): """Perform initial imputation for input X. Parameters ---------- X : ndarray, shape (n_samples, n_features) Input data, where "n_samples" is the number of samples and "n_features" is the number of features. in_fit : bool, default=False Whether function is called in fit. Returns ------- Xt : ndarray, shape (n_samples, n_features) Input data, where "n_samples" is the number of samples and "n_features" is the number of features. X_filled : ndarray, shape (n_samples, n_features) Input data with the most recent imputations. mask_missing_values : ndarray, shape (n_samples, n_features) Input data's missing indicator matrix, where "n_samples" is the number of samples and "n_features" is the number of features. X_missing_mask : ndarray, shape (n_samples, n_features) Input data's mask matrix indicating missing datapoints, where "n_samples" is the number of samples and "n_features" is the number of features. """ if is_scalar_nan(self.missing_values): force_all_finite = "allow-nan" else: force_all_finite = True X = self._validate_data(X, dtype=FLOAT_DTYPES, order="F", reset=in_fit, force_all_finite=force_all_finite) _check_inputs_dtype(X, self.missing_values) X_missing_mask = _get_mask(X, self.missing_values) mask_missing_values = X_missing_mask.copy() if self.initial_imputer_ is None: self.initial_imputer_ = SimpleImputer( missing_values=self.missing_values, strategy=self.initial_strategy ) X_filled = self.initial_imputer_.fit_transform(X) else: X_filled = self.initial_imputer_.transform(X) valid_mask = np.flatnonzero(np.logical_not( np.isnan(self.initial_imputer_.statistics_))) Xt = X[:, valid_mask] mask_missing_values = mask_missing_values[:, valid_mask] return Xt, X_filled, mask_missing_values, X_missing_mask @staticmethod def _validate_limit(limit, limit_type, n_features): """Validate the limits (min/max) of the feature values Converts scalar min/max limits to vectors of shape (n_features,) Parameters ---------- limit: scalar or array-like The user-specified limit (i.e, min_value or max_value) limit_type: string, "max" or "min" n_features: Number of features in the dataset Returns ------- limit: ndarray, shape(n_features,) Array of limits, one for each feature """ limit_bound = np.inf if limit_type == "max" else -np.inf limit = limit_bound if limit is None else limit if np.isscalar(limit): limit = np.full(n_features, limit) limit = check_array( limit, force_all_finite=False, copy=False, ensure_2d=False ) if not limit.shape[0] == n_features: raise ValueError( f"'{limit_type}_value' should be of " f"shape ({n_features},) when an array-like " f"is provided. Got {limit.shape}, instead." ) return limit def fit_transform(self, X, y=None): """Fits the imputer on X and return the transformed X. Parameters ---------- X : array-like, shape (n_samples, n_features) Input data, where "n_samples" is the number of samples and "n_features" is the number of features. y : ignored. Returns ------- Xt : array-like, shape (n_samples, n_features) The imputed input data. """ self.random_state_ = getattr(self, "random_state_", check_random_state(self.random_state)) if self.max_iter < 0: raise ValueError( "'max_iter' should be a positive integer. Got {} instead." .format(self.max_iter)) if self.tol < 0: raise ValueError( "'tol' should be a non-negative float. Got {} instead." .format(self.tol) ) if self.estimator is None: from ..linear_model import BayesianRidge self._estimator = BayesianRidge() else: self._estimator = clone(self.estimator) self.imputation_sequence_ = [] self.initial_imputer_ = None X, Xt, mask_missing_values, complete_mask = ( self._initial_imputation(X, in_fit=True)) super()._fit_indicator(complete_mask) X_indicator = super()._transform_indicator(complete_mask) if self.max_iter == 0 or np.all(mask_missing_values): self.n_iter_ = 0 return super()._concatenate_indicator(Xt, X_indicator) # Edge case: a single feature. We return the initial ... if Xt.shape[1] == 1: self.n_iter_ = 0 return super()._concatenate_indicator(Xt, X_indicator) self._min_value = self._validate_limit( self.min_value, "min", X.shape[1]) self._max_value = self._validate_limit( self.max_value, "max", X.shape[1]) if not np.all(np.greater(self._max_value, self._min_value)): raise ValueError( "One (or more) features have min_value >= max_value.") # order in which to impute # note this is probably too slow for large feature data (d > 100000) # and a better way would be good. # see: https://goo.gl/KyCNwj and subsequent comments ordered_idx = self._get_ordered_idx(mask_missing_values) self.n_features_with_missing_ = len(ordered_idx) abs_corr_mat = self._get_abs_corr_mat(Xt) n_samples, n_features = Xt.shape if self.verbose > 0: print("[IterativeImputer] Completing matrix with shape %s" % (X.shape,)) start_t = time() if not self.sample_posterior: Xt_previous = Xt.copy() normalized_tol = self.tol * np.max( np.abs(X[~mask_missing_values]) ) for self.n_iter_ in range(1, self.max_iter + 1): if self.imputation_order == 'random': ordered_idx = self._get_ordered_idx(mask_missing_values) for feat_idx in ordered_idx: neighbor_feat_idx = self._get_neighbor_feat_idx(n_features, feat_idx, abs_corr_mat) Xt, estimator = self._impute_one_feature( Xt, mask_missing_values, feat_idx, neighbor_feat_idx, estimator=None, fit_mode=True) estimator_triplet = _ImputerTriplet(feat_idx, neighbor_feat_idx, estimator) self.imputation_sequence_.append(estimator_triplet) if self.verbose > 1: print('[IterativeImputer] Ending imputation round ' '%d/%d, elapsed time %0.2f' % (self.n_iter_, self.max_iter, time() - start_t)) if not self.sample_posterior: inf_norm = np.linalg.norm(Xt - Xt_previous, ord=np.inf, axis=None) if self.verbose > 0: print('[IterativeImputer] ' 'Change: {}, scaled tolerance: {} '.format( inf_norm, normalized_tol)) if inf_norm < normalized_tol: if self.verbose > 0: print('[IterativeImputer] Early stopping criterion ' 'reached.') break Xt_previous = Xt.copy() else: if not self.sample_posterior: warnings.warn("[IterativeImputer] Early stopping criterion not" " reached.", ConvergenceWarning) Xt[~mask_missing_values] = X[~mask_missing_values] return super()._concatenate_indicator(Xt, X_indicator) def transform(self, X): """Imputes all missing values in X. Note that this is stochastic, and that if random_state is not fixed, repeated calls, or permuted input, will yield different results. Parameters ---------- X : array-like of shape (n_samples, n_features) The input data to complete. Returns ------- Xt : array-like, shape (n_samples, n_features) The imputed input data. """ check_is_fitted(self) X, Xt, mask_missing_values, complete_mask = self._initial_imputation(X) X_indicator = super()._transform_indicator(complete_mask) if self.n_iter_ == 0 or np.all(mask_missing_values): return super()._concatenate_indicator(Xt, X_indicator) imputations_per_round = len(self.imputation_sequence_) // self.n_iter_ i_rnd = 0 if self.verbose > 0: print("[IterativeImputer] Completing matrix with shape %s" % (X.shape,)) start_t = time() for it, estimator_triplet in enumerate(self.imputation_sequence_): Xt, _ = self._impute_one_feature( Xt, mask_missing_values, estimator_triplet.feat_idx, estimator_triplet.neighbor_feat_idx, estimator=estimator_triplet.estimator, fit_mode=False ) if not (it + 1) % imputations_per_round: if self.verbose > 1: print('[IterativeImputer] Ending imputation round ' '%d/%d, elapsed time %0.2f' % (i_rnd + 1, self.n_iter_, time() - start_t)) i_rnd += 1 Xt[~mask_missing_values] = X[~mask_missing_values] return super()._concatenate_indicator(Xt, X_indicator) def fit(self, X, y=None): """Fits the imputer on X and return self. Parameters ---------- X : array-like, shape (n_samples, n_features) Input data, where "n_samples" is the number of samples and "n_features" is the number of features. y : ignored Returns ------- self : object Returns self. """ self.fit_transform(X) return self