simple_imputer.py

# Copyright 2018 Amazon.com, Inc. or its affiliates. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"). You may not
# use this file except in compliance with the License. A copy of the License
# is located at
#
#     http://aws.amazon.com/apache2.0/
#
# or in the "license" file accompanying this file. This file is distributed on
# an "AS IS" BASIS, WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either
# express or implied. See the License for the specific language governing
# permissions and limitations under the License.

"""

DataWig SimpleImputer:
Uses some simple default encoders and featurizers that usually yield decent imputation quality

"""
import glob
import inspect
import json
import os
import pickle
import shutil
from typing import List, Dict, Any, Callable

import mxnet as mx
import pandas as pd
import numpy as np
from pandas.api.types import is_numeric_dtype

from ._hpo import _HPO
from .column_encoders import BowEncoder, CategoricalEncoder, NumericalEncoder, TfIdfEncoder
from .imputer import Imputer
from .mxnet_input_symbols import BowFeaturizer, NumericalFeaturizer
from .utils import logger, get_context, random_split, rand_string, flatten_dict, merge_dicts, set_stream_log_level
from .imputer import Imputer
from .iterators import INSTANCE_WEIGHT_COLUMN



class SimpleImputer:
    """

    SimpleImputer model based on n-grams of concatenated strings of input columns and concatenated
    numerical features, if provided.

    Given a data frame with string columns, a model is trained to predict observed values in label
    column using values observed in other columns.

    The model can then be used to impute missing values.

    :param input_columns: list of input column names (as strings)
    :param output_column: output column name (as string)
    :param output_path: path to store model and metrics
    :param num_hash_buckets: number of hash buckets used for the n-gram hashing vectorizer, only
                                used for non-numerical input columns, ignored otherwise
    :param num_labels: number of imputable values considered after, only used for non-numerical
                        input columns, ignored otherwise
    :param tokens: string, 'chars' or 'words' (default 'chars'), determines tokenization strategy
                for n-grams, only used for non-numerical input columns, ignored otherwise
    :param numeric_latent_dim: int, number of latent dimensions for hidden layer of NumericalFeaturizers;
                only used for numerical input columns, ignored otherwise
    :param numeric_hidden_layers: number of numeric hidden layers
    :param is_explainable: if this is True, a stateful tf-idf encoder is used that allows
                           explaining classes and single instances


    Example usage:


    from datawig.simple_imputer import SimpleImputer
    import pandas as pd

    fn_train = os.path.join(datawig_test_path, "resources", "shoes", "train.csv.gz")
    fn_test = os.path.join(datawig_test_path, "resources", "shoes", "test.csv.gz")

    df_train = pd.read_csv(training_data_files)
    df_test = pd.read_csv(testing_data_files)

    output_path = "imputer_model"

    # set up imputer model
    imputer = SimpleImputer( input_columns=['item_name', 'bullet_point'], output_column='brand')

    # train the imputer model
    imputer = imputer.fit(df_train)

    # obtain imputations
    imputations = imputer.predict(df_test)

    """

    def __init__(self,
                 input_columns: List[str],
                 output_column: str,
                 output_path: str = "",
                 num_hash_buckets: int = int(2 ** 15),
                 num_labels: int = 100,
                 tokens: str = 'chars',
                 numeric_latent_dim: int = 100,
                 numeric_hidden_layers: int = 1,
                 is_explainable: bool = False) -> None:

        for col in input_columns:
            if not isinstance(col, str):
                raise ValueError("SimpleImputer.input_columns must be str type, was {}".format(type(col)))

        if not isinstance(output_column, str):
            raise ValueError("SimpleImputer.output_column must be str type, was {}".format(type(output_column)))

        self.input_columns = input_columns
        self.output_column = output_column
        self.num_hash_buckets = num_hash_buckets
        self.num_labels = num_labels
        self.tokens = tokens
        self.numeric_latent_dim = numeric_latent_dim
        self.numeric_hidden_layers = numeric_hidden_layers
        self.output_path = output_path
        self.imputer = None
        self.hpo = None
        self.numeric_columns = []
        self.string_columns = []
        self.hpo = _HPO()
        self.is_explainable = is_explainable

    def check_data_types(self, data_frame: pd.DataFrame) -> None:
        """

        Checks whether a column contains string or numeric data

        :param data_frame:
        :return:
        """
        self.numeric_columns = [c for c in self.input_columns if is_numeric_dtype(data_frame[c])]
        self.string_columns = list(set(self.input_columns) - set(self.numeric_columns))
        self.output_type = 'numeric' if is_numeric_dtype(data_frame[self.output_column]) else 'string'

        logger.debug(
            "Assuming {} numeric input columns: {}".format(len(self.numeric_columns),
                                                           ", ".join(self.numeric_columns)))
        logger.debug("Assuming {} string input columns: {}".format(len(self.string_columns),
                                                                  ", ".join(self.string_columns)))

    @staticmethod
    def _is_categorical(col: pd.Series,
                        n_samples: int = 100,
                        max_unique_fraction=0.05) -> bool:
        """

        A heuristic to check whether a column is categorical:
        a column is considered categorical (as opposed to a plain text column)
        if the relative cardinality is max_unique_fraction or less.

        :param col: pandas Series containing strings
        :param n_samples: number of samples used for heuristic (default: 100)
        :param max_unique_fraction: maximum relative cardinality.

        :return: True if the column is categorical according to the heuristic

        """

        sample = col.sample(n=n_samples, replace=len(col) < n_samples).unique()

        return sample.shape[0] / n_samples < max_unique_fraction

    def fit_hpo(self,
                train_df: pd.DataFrame,
                test_df: pd.DataFrame = None,
                hps: dict = None,
                num_evals: int = 10,
                max_running_hours: float = 96.0,
                hpo_run_name: str = None,
                user_defined_scores: list = None,
                num_epochs: int = None,
                patience: int = None,
                test_split: float = .2,
                weight_decay: List[float] = None,
                batch_size: int = 16,
                num_hash_bucket_candidates: List[float] = [2 ** exp for exp in [12, 15, 18]],
                tokens_candidates: List[str] = ['words', 'chars'],
                numeric_latent_dim_candidates: List[int] = None,
                numeric_hidden_layers_candidates: List[int] = None,
                final_fc_hidden_units: List[List[int]] = None,
                learning_rate_candidates: List[float] = None,
                normalize_numeric: bool = True,
                hpo_max_train_samples: int = None,
                ctx: mx.context = get_context()) -> Any:

        """
        Fits an imputer model with hyperparameter optimization. The parameter ranges are searched randomly.

        Grids are specified using the *_candidates arguments (old)
        or with more flexibility via the dictionary hps.

        :param train_df: training data as dataframe
        :param test_df: test data as dataframe; if not provided, a ratio of test_split of the
            training data are used as test data
        :param hps: nested dictionary where hps[global][parameter_name] is list of parameters. Similarly,
            hps[column_name][parameter_name] is a list of parameter values for each input column.
            Further, hps[column_name]['type'] is in ['numeric', 'categorical', 'string'] and is
            inferred if not provided.
        :param num_evals: number of evaluations for random search
        :param max_running_hours: Time before the hpo run is terminated in hours.
        :param hpo_run_name: string to identify the current hpo run.
        :param user_defined_scores: list with entries (Callable, str), where callable is a function
            accepting **kwargs true, predicted, confidence. Allows custom scoring functions.

        Below are parameters of the old implementation, kept to ascertain backwards compatibility.
        :param num_epochs: maximal number of training epochs (default 10)
        :param patience: used for early stopping; after [patience] epochs with no improvement,
            training is stopped. (default 3)
        :param test_split: if no test_df is provided this is the ratio of test data to be held
            separate for determining model convergence
        :param weight_decay: regularizer (default 0)
        :param batch_size (default 16)
        :param num_hash_bucket_candidates: candidates for gridsearch hyperparameter
            optimization (default [2**10, 2**13, 2**15, 2**18, 2**20])
        :param tokens_candidates: candidates for tokenization (default ['words', 'chars'])
        :param numeric_latent_dim_candidates: candidates for latent dimensionality of
            numerical features (default [10, 50, 100])
        :param numeric_hidden_layers_candidates: candidates for number of hidden layers of
        :param final_fc_hidden_units: list of lists w/ dimensions for FC layers after the
            final concatenation (NOTE: for HPO, this expects a list of lists)
        :param learning_rate_candidates: learning rate for stochastic gradient descent (default 4e-4)
            numerical features (default [0, 1, 2])
        :param learning_rate_candidates: candidates for learning rate (default [1e-1, 1e-2, 1e-3])
        :param normalize_numeric: boolean indicating whether or not to normalize numeric values
        :param hpo_max_train_samples: training set size for hyperparameter optimization.
            use is deprecated.
        :param ctx: List of mxnet contexts (if no gpu's available, defaults to [mx.cpu()])
            User can also pass in a list gpus to be used, ex. [mx.gpu(0), mx.gpu(2), mx.gpu(4)]
            This parameter is deprecated.

        :return: pd.DataFrame with with hyper-parameter configurations and results
        """

        # generate dictionary with default hyperparameter settings. Overwrite these defaults
        # with configurations that were passed via this functions API wherever applicable.
        default_hps = dict()
        default_hps['global'] = dict()
        if learning_rate_candidates:
            default_hps['global']['learning_rate'] = learning_rate_candidates
        if weight_decay:
            default_hps['global']['weight_decay'] = weight_decay
        if num_epochs:
            default_hps['global']['num_epochs'] = [num_epochs]

        if patience:
            default_hps['global']['patience'] = [patience]

        if batch_size:
            default_hps['global']['batch_size'] = [batch_size]
        if final_fc_hidden_units:
            default_hps['global']['final_fc_hidden_units'] = final_fc_hidden_units

        default_hps['string'] = {}
        if num_hash_bucket_candidates:
            default_hps['string']['max_tokens'] = num_hash_bucket_candidates

        if tokens_candidates:
            default_hps['string']['tokens'] = [[c] for c in tokens_candidates]

        default_hps['categorical'] = {}
        if num_hash_bucket_candidates:
            default_hps['categorical']['max_tokens'] = num_hash_bucket_candidates

        default_hps['numeric'] = {}
        if normalize_numeric:
            default_hps['numeric']['normalize'] = [normalize_numeric]
        if numeric_latent_dim_candidates:
            default_hps['numeric']['numeric_latent_dim'] = numeric_latent_dim_candidates

        if numeric_hidden_layers_candidates:
            default_hps['numeric']['numeric_hidden_layers'] = numeric_hidden_layers_candidates

        if hps is None:
            hps = {}

        # give parameters in `hps` precedence over default parameters
        parameters_in_both = set(default_hps.keys()).intersection(set(hps.keys()))
        for param in parameters_in_both:
            del default_hps[param]
        hps = merge_dicts(hps, default_hps)

        if user_defined_scores is None:
            user_defined_scores = []

        if test_df is None:
            train_df, test_df = random_split(train_df, [1-test_split, test_split])

        self.check_data_types(train_df)  # infer data types, saved self.string_columns, self.numeric_columns
        self.hpo.tune(train_df, test_df, hps, num_evals, max_running_hours, user_defined_scores, hpo_run_name, self)
        self.save()

        return self

    def fit(self,
            train_df: pd.DataFrame,
            test_df: pd.DataFrame = None,
            ctx: mx.context = get_context(),
            learning_rate: float = 4e-3,
            num_epochs: int = 100,
            patience: int = 5,
            test_split: float = .1,
            weight_decay: float = 0.,
            batch_size: int = 16,
            final_fc_hidden_units: List[int] = None,
            calibrate: bool = True,
            class_weights: dict = None,
            instance_weights: list = None) -> Any:
        """

        Trains and stores imputer model

        :param train_df: training data as dataframe
        :param test_df: test data as dataframe; if not provided, a ratio of test_split of the
                            training data are used as test data
        :param ctx: List of mxnet contexts (if no gpu's available, defaults to [mx.cpu()])
                    User can also pass in a list gpus to be used, ex. [mx.gpu(0), mx.gpu(2), mx.gpu(4)]
        :param learning_rate: learning rate for stochastic gradient descent (default 4e-4)
        :param num_epochs: maximal number of training epochs (default 10)
        :param patience: used for early stopping; after [patience] epochs with no improvement,
                            training is stopped. (default 3)
        :param test_split: if no test_df is provided this is the ratio of test data to be held
                            separate for determining model convergence
        :param weight_decay: regularizer (default 0)
        :param batch_size (default 16)
        :param final_fc_hidden_units: list dimensions for FC layers after the final concatenation
        :param calibrate: Control automatic model calibration
        :param class_weights: Dictionary with labels as keys and weights as values.
                              Weighs each instance's contribution to the likelihood based on the corresponding class.
        :param instance_weights: List of weights for each instance in train_df.
        """

        # add weights to training data if provided
        train_df = self.__add_weights_to_df(train_df, class_weights, instance_weights, in_place=False)

        self.check_data_types(train_df)

        data_encoders = []
        data_columns = []

        if len(self.string_columns) > 0:
            string_feature_column = "ngram_features-" + rand_string(10)
            if self.is_explainable:
                data_encoders += [TfIdfEncoder(input_columns=self.string_columns,
                                               output_column=string_feature_column,
                                               max_tokens=self.num_hash_buckets,
                                               tokens=self.tokens)]
            else:
                data_encoders += [BowEncoder(input_columns=self.string_columns,
                                             output_column=string_feature_column,
                                             max_tokens=self.num_hash_buckets,
                                             tokens=self.tokens)]
            data_columns += [
                BowFeaturizer(field_name=string_feature_column, max_tokens=self.num_hash_buckets)]

        if len(self.numeric_columns) > 0:
            numerical_feature_column = "numerical_features-" + rand_string(10)

            data_encoders += [NumericalEncoder(input_columns=self.numeric_columns,
                                               output_column=numerical_feature_column)]

            data_columns += [
                NumericalFeaturizer(field_name=numerical_feature_column, numeric_latent_dim=self.numeric_latent_dim,
                                    numeric_hidden_layers=self.numeric_hidden_layers)]

        if is_numeric_dtype(train_df[self.output_column]):
            label_column = [NumericalEncoder(self.output_column, normalize=True)]
            logger.debug("Assuming numeric output column: {}".format(self.output_column))
        else:
            label_column = [CategoricalEncoder(self.output_column, max_tokens=self.num_labels)]
            logger.debug("Assuming categorical output column: {}".format(self.output_column))

        # to make consecutive calls to .fit() continue where the previous call finished
        if self.imputer is None:
            self.imputer = Imputer(data_encoders=data_encoders,
                                   data_featurizers=data_columns,
                                   label_encoders=label_column,
                                   output_path=self.output_path)

        self.output_path = self.imputer.output_path

        self.imputer = self.imputer.fit(train_df, test_df, ctx, learning_rate, num_epochs, patience,
                                        test_split,
                                        weight_decay, batch_size,
                                        final_fc_hidden_units=final_fc_hidden_units,
                                        calibrate=calibrate)
        self.save()

        return self

    def predict(self,
                data_frame: pd.DataFrame,
                precision_threshold: float = 0.0,
                imputation_suffix: str = "_imputed",
                score_suffix: str = "_imputed_proba",
                inplace: bool = False):
        """
        Imputes most likely value if it is above a certain precision threshold determined on the
            validation set
        Precision is calculated as part of the `datawig.evaluate_and_persist_metrics` function.

        Returns original dataframe with imputations and respective likelihoods as estimated by
        imputation model; in additional columns; names of imputation columns are that of the label
        suffixed with `imputation_suffix`, names of respective likelihood columns are suffixed
        with `score_suffix`

        :param data_frame:   data frame (pandas)
        :param precision_threshold: double between 0 and 1 indicating precision threshold
        :param imputation_suffix: suffix for imputation columns
        :param score_suffix: suffix for imputation score columns
        :param inplace: add column with imputed values and column with confidence scores to data_frame, returns the
            modified object (True). Create copy of data_frame with additional columns, leave input unmodified (False).
        :return: data_frame original dataframe with imputations and likelihood in additional column
        """
        imputations = self.imputer.predict(data_frame, precision_threshold, imputation_suffix,
                                           score_suffix, inplace=inplace)

        return imputations

    def explain(self, label: str, k: int = 10, label_column: str = None) -> dict:
        """
        Return dictionary with a list of tuples for each explainable input column.
        Each tuple denotes one of the top k features with highest correlation to the label.

        :param label: label value to explain
        :param k: number of explanations for each input encoder to return. If not given, return top 10 explanations.
        :param label_column: name of label column to be explained (optional, defaults to the first available column.)
        """
        if self.imputer is None:
            raise ValueError("Need to call .fit() before")

        return self.imputer.explain(label, k, label_column)

    def explain_instance(self,
                         instance: pd.core.series.Series,
                         k: int = 10,
                         label_column: str = None,
                         label: str = None) -> dict:
        """
        Return dictionary with list of tuples for each explainable input column of the given instance.
        Each entry shows the most highly correlated features to the given label
        (or the top predicted label of not provided).

        :param instance: row of data frame (or dictionary)
        :param k: number of explanations (ngrams) for text inputs
        :param label_column: name of label column to be explained (optional)
        :param label: explain why instance is classified as label, otherwise explain top-label per input
        """
        if self.imputer is None:
            raise ValueError("Need to call .fit() before")

        return self.imputer.explain_instance(instance, k, label_column, label)

    @staticmethod
    def complete(data_frame: pd.DataFrame,
                 precision_threshold: float = 0.0,
                 inplace: bool = False,
                 hpo: bool = False,
                 verbose: int = 0,
                 num_epochs: int = 100,
                 iterations: int = 1,
                 output_path: str = "."):
        """
        Given a dataframe with missing values, this function detects all imputable columns, trains an imputation model
        on all other columns and imputes values for each missing value.
        Several imputation iterators can be run.
        Imputable columns are either numeric columns or non-numeric categorical columns; for determining whether a
            column is categorical (as opposed to a plain text column) we use the following heuristic:
            a non-numeric categorical column should have least 10 times as many rows as there were unique values
        If an imputation model did not reach the precision specified in the precision_threshold parameter for a given
            imputation value, that value will not be imputed; thus depending on the precision_threshold, the returned
            dataframe can still contain some missing values.
        For numeric columns, we do not filter for accuracy.
        :param data_frame: original dataframe
        :param precision_threshold: precision threshold for categorical imputations (default: 0.0)
        :param inplace: whether or not to perform imputations inplace (default: False)
        :param hpo: whether or not to perform hyperparameter optimization (default: False)
        :param verbose: verbosity level, values > 0 log to stdout (default: 0)
        :param num_epochs: number of epochs for each imputation model training (default: 100)
        :param iterations: number of iterations for iterative imputation (default: 1)
        :param output_path: path to store model and metrics
        :return: dataframe with imputations
        """

        # TODO: should we expose temporary dir for model serialization to avoid crashes due to not-writable dirs?
        
        missing_mask = data_frame.copy().isnull()

        if inplace is False:
            data_frame = data_frame.copy()

        if verbose == 0:
            set_stream_log_level("ERROR")

        numeric_columns = [c for c in data_frame.columns if is_numeric_dtype(data_frame[c])]
        string_columns = list(set(data_frame.columns) - set(numeric_columns))
        logger.debug("Assuming numerical columns: {}".format(", ".join(numeric_columns)))

        col_set = set(numeric_columns + string_columns)

        categorical_columns = [col for col in string_columns if SimpleImputer._is_categorical(data_frame[col])]
        logger.debug("Assuming categorical columns: {}".format(", ".join(categorical_columns)))
        for _ in range(iterations):
            for output_col in set(numeric_columns) | set(categorical_columns):
                # train on all input columns but the to-be-imputed one
                input_cols = list(col_set - set([output_col]))

                # train on all observed values
                idx_missing = missing_mask[output_col]

                imputer = SimpleImputer(input_columns=input_cols,
                                        output_column=output_col,
                                        output_path=os.path.join(output_path, output_col))
                if hpo:
                    imputer.fit_hpo(data_frame.loc[~idx_missing, :],
                                    patience=5 if output_col in categorical_columns else 20,
                                    num_epochs=100,
                                    final_fc_hidden_units=[[0], [10], [50], [100]])
                else:
                    imputer.fit(data_frame.loc[~idx_missing, :],
                                patience=5 if output_col in categorical_columns else 20,
                                num_epochs=num_epochs,
                                calibrate=False)

                tmp = imputer.predict(data_frame, precision_threshold=precision_threshold)
                data_frame.loc[idx_missing, output_col] = tmp[output_col + "_imputed"]

                # remove the directory with logfiles for this column 
                shutil.rmtree(os.path.join(output_path, output_col))
                
                
        return data_frame

    def save(self):
        """

        Saves model to disk; mxnet module and imputer are stored separately

        """
        self.imputer.save()
        simple_imputer_params = {k: v for k, v in self.__dict__.items() if k not in ['imputer']}
        pickle.dump(simple_imputer_params,
                    open(os.path.join(self.output_path, "simple_imputer.pickle"), "wb"))

    def load_metrics(self) -> Dict[str, Any]:
        """
        Loads various metrics of the internal imputer model, returned as dictionary
        :return: Dict[str,Any]
        """
        assert os.path.isfile(self.imputer.metrics_path), "Metrics File {} not found".format(
            self.imputer.metrics_path)
        metrics = json.load(open(self.imputer.metrics_path))[self.output_column]
        return metrics

    @staticmethod
    def load(output_path: str) -> Any:
        """

        Loads model from output path

        :param output_path: output_path field of trained SimpleImputer model
        :return: SimpleImputer model

        """

        logger.debug("Output path for loading Imputer {}".format(output_path))
        # load pickled model
        simple_imputer_params = pickle.load(
            open(os.path.join(output_path, "simple_imputer.pickle"), "rb"))

        # get constructor args
        constructor_args = inspect.getfullargspec(SimpleImputer.__init__)[0]
        constructor_args = [arg for arg in constructor_args if arg != 'self']

        # get those params that are needed for __init__
        constructor_params = {k: v for k, v in simple_imputer_params.items() if
                              k in constructor_args}
        # instantiate SimpleImputer
        simple_imputer = SimpleImputer(**constructor_params)
        # set all other fields
        for key, value in simple_imputer_params.items():
            if key not in constructor_args:
                setattr(simple_imputer, key, value)

        # set imputer model, only if it exists.
        simple_imputer.imputer = Imputer.load(output_path)

        return simple_imputer

    def load_hpo_model(self, hpo_name: int = None):
        """
        Load model after hyperparameter optimisation has ran. Overwrites local artifacts of self.imputer.

        :param hpo_name: Index of the model to be loaded. Default,
                    load model with highest weighted precision or mean squared error.

        :return: imputer object
        """

        if self.hpo.results is None:
            logger.warning('No hpo run available. Run hpo by calling SimpleImputer.fit_hpo().')
            return

        if hpo_name is None:
            if self.output_type == 'numeric':
                hpo_name = self.hpo.results['mse'].astype(float).idxmin()
                logger.debug("Selecting imputer with minimal mean squared error.")
            else:
                hpo_name = self.hpo.results['precision_weighted'].astype(float).idxmax()
                logger.debug("Selecting imputer with maximal weighted precision.")

        # copy artifacts from hp run to self.output_path
        imputer_dir = self.output_path + str(hpo_name)
        files_to_copy = glob.glob(imputer_dir + '/*.json') + \
                        glob.glob(imputer_dir + '/*.pickle') + \
                        glob.glob(imputer_dir + '/*.params')
        for file_name in files_to_copy:
            shutil.copy(file_name,  self.output_path)

        self.imputer = Imputer.load(self.output_path)

    def __deserialize_confusion_matrix(self) -> np.ndarray:
        """
        Return normalized confusion matrix for trained simple_imputer
        """

        # labels need to be sorted consistently for computation of confusion matrices etc.
        labels = sorted(self.imputer.label_encoders[0].idx_to_token.values())

        # invert normalized empirical confusion matrix for test data
        ids = dict((label, index) for (index, label) in enumerate(labels))
        confusion_matrix_from_file = self.load_metrics()['confusion_matrix']
        confusion_test = np.empty([len(confusion_matrix_from_file), len(confusion_matrix_from_file)])

        # make matrix from nested dictionary that stores the confusion matrix
        for row in confusion_matrix_from_file:
            for entry in row:
                if type(entry) is str:
                    row_idx = ids[entry]
                else:
                    for col in entry:
                        col_idx = ids[col[0]]
                        confusion_test[row_idx, col_idx] = col[1]

        return confusion_test / confusion_test.sum()  # normalize

    def check_for_label_shift(self, target_data: pd.DataFrame) -> dict:

        """
        Detect label shift in the validation data

        :param test_data: data frame that contains labels
        :param target_data: unlabelled data for which predictions are to be generated

        :return: dictionary with labels as keys and weights as values.
        """

        # labels need to be sorted consistently for computation of confusion matrices etc.
        labels = sorted(self.imputer.label_encoders[0].idx_to_token.values())

        imputed_target = self.predict(target_data)  # generate predictions for test and target data
        confusion_test = self.__deserialize_confusion_matrix()

        # compute estimates of label marginals for test and target data
        marginals_test = pd.Series(confusion_test.sum(axis=0), index=labels)
        marginals_target = imputed_target[self.output_column + '_imputed'].value_counts(
            normalize=True, sort=False)[labels]

        # estimate the ratio of marginals and store as dictionary.
        label_weights = np.linalg.solve(confusion_test, marginals_target)
        label_weights_dict = dict((label, max(weight, 0)) for label, weight in zip(labels, label_weights))

        # estimate marginals of true labels
        true_marginals_target = np.diag(marginals_test).dot(label_weights)

        logger.warning('\n\tThe estimated label marginals are ' + str(list(zip(labels, true_marginals_target))) +
                    '\n\tMarginals in the training data are ' + str(list(zip(labels, marginals_test))) +
                    '\n\tReweighing factors for empirical risk minimization' + str(label_weights_dict))

        if np.any(marginals_test < 0):
            logger.warning('\n\tEstimated label marginals are invalid. Proceed with caution.')

        return label_weights_dict

    def __add_weights_to_df(self,
                            df: pd.DataFrame,
                            class_weights: dict = None,
                            instance_weights: list = None,
                            in_place: bool = True) -> pd.DataFrame:
        """
        Add additional column to data frame inplace, with entries provided either
        (1) as dict values in class_dictionary with rows selected on dict keys in weights dictionary
            in correspondence to label_column.
        (2) or as list with weights for each instance

        :param df: input data
        :param class_weights: dictionary with label names and corresponding weights
        :param instance_weights: list of weights for each instance
        :param label_column: name of label column
        :param in_place: If true, append column to df, otherwise create copy.
        """

        # Nothing to do if no weights have been passed.
        if class_weights is None and instance_weights is None:
            return df

        # Check that only one type of weights is provided
        assert (class_weights is None or instance_weights is None), \
            "Please provide class_weights XOR instance_weights."

        if in_place is False:
            df_new = df.copy()
        else:
            df_new = df

        if class_weights is not None:
            df_new[INSTANCE_WEIGHT_COLUMN] = 0.0
            for label, weight in class_weights.items():
                df_new.loc[df_new[self.output_column] == label, INSTANCE_WEIGHT_COLUMN] = weight

        elif instance_weights is not None:
            df_new[INSTANCE_WEIGHT_COLUMN] = instance_weights

        return df_new