hc99's picture
Add files using upload-large-folder tool
fc0f7bd verified
# Copyright (c) Microsoft Corporation. All rights reserved.
# Licensed under the MIT License.
"""Threshold Optimization Post Processing algorithm.
This is based on M. Hardt, E. Price, N. Srebro's paper
"`Equality of Opportunity in Supervised Learning
<https://arxiv.org/pdf/1610.02413.pdf>`_" for binary
classification with one categorical sensitive feature.
"""
import logging
import numpy as np
import pandas as pd
import random
from sklearn.exceptions import NotFittedError
from fairlearn.postprocessing import PostProcessing
from ._constants import (LABEL_KEY, SCORE_KEY, SENSITIVE_FEATURE_KEY, OUTPUT_SEPARATOR,
DEMOGRAPHIC_PARITY, EQUALIZED_ODDS)
from ._roc_curve_utilities import _interpolate_curve, _get_roc
from ._interpolated_prediction import InterpolatedPredictor
# various error messages
DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE = "{} need to be of equal length."
EMPTY_INPUT_ERROR_MESSAGE = "At least one of sensitive_features, labels, or scores are empty."
NON_BINARY_LABELS_ERROR_MESSAGE = "Labels other than 0/1 were provided."
INPUT_DATA_FORMAT_ERROR_MESSAGE = "The only allowed input data formats are: " \
"list, numpy.ndarray, pandas.DataFrame, pandas.Series. " \
"Your provided data was of types ({}, {}, {})"
NOT_SUPPORTED_CONSTRAINTS_ERROR_MESSAGE = "Currently only {} and {} are supported " \
"constraints.".format(DEMOGRAPHIC_PARITY, EQUALIZED_ODDS)
PREDICT_BEFORE_FIT_ERROR_MESSAGE = "It is required to call 'fit' before 'predict'."
MULTIPLE_DATA_COLUMNS_ERROR_MESSAGE = "Post processing currently only supports a single " \
"column in {}."
SENSITIVE_FEATURE_NAME_CONFLICT_DETECTED_ERROR_MESSAGE = "A sensitive feature named {} or {} " \
"was detected. Please rename your column and try again.".format(SCORE_KEY, LABEL_KEY)
SCORES_DATA_TOO_MANY_COLUMNS_ERROR_MESSAGE = "The provided scores data contains multiple columns."
UNEXPECTED_DATA_TYPE_ERROR_MESSAGE = "Unexpected data type {} encountered."
_SUPPORTED_CONSTRAINTS = [DEMOGRAPHIC_PARITY, EQUALIZED_ODDS]
logger = logging.getLogger(__name__)
class ThresholdOptimizer(PostProcessing):
"""An Estimator based on the threshold optimization approach.
The procedure followed is described in detail in
`Hardt et al. (2016) <https://arxiv.org/abs/1610.02413>`_.
:param unconstrained_predictor: The trained predictor whose output will be post processed
:type unconstrained_predictor: A trained predictor
:param estimator: An untrained estimator that will be trained, and
subsequently its output will be post processed
:type estimator: An untrained estimator
:param grid_size: The number of ticks on the grid over which we evaluate the curves.
A large grid_size means that we approximate the actual curve, so it increases the chance
of being very close to the actual best solution.
:type grid_size: int
:param flip: Allow flipping to negative weights if it improves accuracy.
:type flip: bool
:param plot: Show ROC/selection-error plot if True
:type plot: bool
"""
def __init__(self, *, unconstrained_predictor=None, estimator=None,
constraints=DEMOGRAPHIC_PARITY, grid_size=1000, flip=True, plot=False):
super(ThresholdOptimizer, self).__init__(
unconstrained_predictor=unconstrained_predictor,
estimator=estimator,
constraints=constraints)
self._constraints = constraints
if self._constraints not in _SUPPORTED_CONSTRAINTS:
raise ValueError(NOT_SUPPORTED_CONSTRAINTS_ERROR_MESSAGE)
self._grid_size = grid_size
self._flip = flip
self._plot = plot
self._post_processed_predictor_by_sensitive_feature = None
def fit(self, X, y, *, sensitive_features, **kwargs):
"""Fit the model.
The fit is based on training features and labels, sensitive features,
as well as the fairness-unaware predictor or estimator. If an estimator was passed
in the constructor this fit method will call `fit(X, y, **kwargs)` on said estimator.
:param X: The feature matrix
:type X: numpy.ndarray or pandas.DataFrame
:param y: The label vector
:type y: numpy.ndarray, pandas.DataFrame, pandas.Series, or list
:param sensitive_features: sensitive features to identify groups by, currently allows
only a single column
:type sensitive_features: currently 1D array as numpy.ndarray, list, pandas.DataFrame,
or pandas.Series
"""
self._validate_input_data(X, sensitive_features, y)
# postprocessing can't handle 0/1 as floating point numbers, so this converts it to int
if type(y) in [np.ndarray, pd.DataFrame, pd.Series]:
y = y.astype(int)
else:
y = [int(y_val) for y_val in y]
if self._estimator:
# train estimator on data first
self._validate_estimator()
self._estimator.fit(X, y, **kwargs)
self._unconstrained_predictor = self._estimator
self._validate_predictor()
scores = self._unconstrained_predictor.predict(X)
threshold_optimization_method = None
if self._constraints == DEMOGRAPHIC_PARITY:
threshold_optimization_method = \
_threshold_optimization_demographic_parity
elif self._constraints == EQUALIZED_ODDS:
threshold_optimization_method = \
_threshold_optimization_equalized_odds
else:
raise ValueError(NOT_SUPPORTED_CONSTRAINTS_ERROR_MESSAGE)
self._post_processed_predictor_by_sensitive_feature = threshold_optimization_method(
sensitive_features, y, scores, self._grid_size, self._flip, self._plot)
def predict(self, X, *, sensitive_features, random_state=None):
"""Predict label for each sample in X while taking into account sensitive features.
:param X: feature matrix
:type X: numpy.ndarray or pandas.DataFrame
:param sensitive_features: sensitive features to identify groups by, currently allows
only a single column
:type sensitive_features: currently 1D array as numpy.ndarray, list, pandas.DataFrame,
or pandas.Series
:param random_state: set to a constant for reproducibility
:type random_state: int
:return: predictions in numpy.ndarray
"""
if random_state:
random.seed(random_state)
self._validate_post_processed_predictor_is_fitted()
self._validate_input_data(X, sensitive_features)
unconstrained_predictions = self._unconstrained_predictor.predict(X)
positive_probs = _vectorized_prediction(
self._post_processed_predictor_by_sensitive_feature,
sensitive_features,
unconstrained_predictions)
return (positive_probs >= np.random.rand(len(positive_probs))) * 1
def _pmf_predict(self, X, *, sensitive_features):
"""Probabilistic mass function.
:param X: Feature matrix
:type X: numpy.ndarray or pandas.DataFrame
:param sensitive_features: Sensitive features to identify groups by, currently allows
only a single column
:type sensitive_features: Currently 1D array as numpy.ndarray, list, pandas.DataFrame,
or pandas.Series
:return: array of tuples with probabilities for predicting 0 or 1, respectively. The sum
of the two numbers in each tuple needs to add up to 1.
:rtype: numpy.ndarray
"""
self._validate_post_processed_predictor_is_fitted()
self._validate_input_data(X, sensitive_features)
positive_probs = _vectorized_prediction(
self._post_processed_predictor_by_sensitive_feature, sensitive_features,
self._unconstrained_predictor.predict(X))
return np.array([[1.0 - p, p] for p in positive_probs])
def _validate_post_processed_predictor_is_fitted(self):
if not self._post_processed_predictor_by_sensitive_feature:
raise NotFittedError(PREDICT_BEFORE_FIT_ERROR_MESSAGE)
def _validate_input_data(self, X, sensitive_features, y=None):
allowed_input_types = [list, np.ndarray, pd.DataFrame, pd.Series]
if type(X) not in allowed_input_types or \
type(sensitive_features) not in allowed_input_types or \
(y is not None and type(y) not in allowed_input_types):
raise TypeError(INPUT_DATA_FORMAT_ERROR_MESSAGE
.format(type(X).__name__,
type(y).__name__,
type(sensitive_features).__name__))
if len(X) == 0 or len(sensitive_features) == 0 or (y is not None and len(y) == 0):
raise ValueError(EMPTY_INPUT_ERROR_MESSAGE)
if y is None:
if len(X) != len(sensitive_features) or (y is not None and len(X) != len(y)):
raise ValueError(DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE
.format("X and sensitive_features"))
else:
if len(X) != len(sensitive_features) or (y is not None and len(X) != len(y)):
raise ValueError(DIFFERENT_INPUT_LENGTH_ERROR_MESSAGE
.format("X, sensitive_features, and y"))
if set(np.unique(y)) > set([0, 1]):
raise ValueError(NON_BINARY_LABELS_ERROR_MESSAGE)
def _threshold_optimization_demographic_parity(sensitive_features, labels, scores, grid_size=1000,
flip=True, plot=False):
"""Calculate the selection and error rates for every sensitive feature value.
These calculations are made at different
thresholds over the scores. Subsequently weighs each sensitive feature value's error by the
frequency of the sensitive feature value in the data. The minimum error point is the selected
solution, which is recreated by interpolating between two points on the convex hull of all
solutions. Each sensitive feature value has its own predictor in the resulting postprocessed
predictor, which requires the sensitive feature value as an input.
This method assumes that sensitive_features, labels, and scores are non-empty data structures
of equal length, and labels contains only binary labels 0 and 1.
:param sensitive_features: the feature data that determines the groups for which the parity
constraints are enforced
:type sensitive_features: list, numpy.ndarray, pandas.DataFrame, or pandas.Series
:param labels: the labels of the dataset
:type labels: list, numpy.ndarray, pandas.DataFrame, or pandas.Series
:param scores: the scores produced by a predictor's prediction
:type scores: list, numpy.ndarray, pandas.DataFrame, or pandas.Series
:param grid_size: The number of ticks on the grid over which we evaluate the curves.
A large grid_size means that we approximate the actual curve, so it increases the chance
of being very close to the actual best solution.
:type grid_size: int
:param flip: allow flipping to negative weights if it improves accuracy.
:type flip: bool
:param plot: show selection-error plot if True
:type plot: bool
:return: the postprocessed predictor as a function taking the sensitive feature value
and the fairness unaware predictor's score as arguments to produce predictions
"""
n = len(labels)
selection_error_curve = {}
x_grid = np.linspace(0, 1, grid_size + 1)
error_given_selection = 0 * x_grid
data_grouped_by_sensitive_feature = _reformat_and_group_data(
sensitive_features, labels, scores)
for sensitive_feature_value, group in data_grouped_by_sensitive_feature:
# determine probability of current sensitive feature group based on data
n_group = len(group)
n_positive = sum(group[LABEL_KEY])
n_negative = n_group - n_positive
p_sensitive_feature_value = n_group / n
roc_convex_hull = _get_roc(group, sensitive_feature_value, flip=flip)
fraction_negative_label_positive_sample = (
n_negative / n_group) * roc_convex_hull['x']
fraction_positive_label_positive_sample = (
n_positive / n_group) * roc_convex_hull['y']
# Calculate selection to represent the proportion of positive predictions.
roc_convex_hull['selection'] = fraction_negative_label_positive_sample + \
fraction_positive_label_positive_sample
fraction_positive_label_negative_sample = \
(n_positive / n_group) * (1 - roc_convex_hull['y'])
roc_convex_hull['error'] = fraction_negative_label_positive_sample + \
fraction_positive_label_negative_sample
selection_error_curve[sensitive_feature_value] = \
_interpolate_curve(roc_convex_hull, 'selection', 'error', 'operation', x_grid)
# Add up errors for the current group multiplied by the probability of the current group.
# This will help us in identifying the minimum overall error.
error_given_selection += p_sensitive_feature_value * \
selection_error_curve[sensitive_feature_value]['error']
logger.debug(OUTPUT_SEPARATOR)
logger.debug("Processing %s", str(sensitive_feature_value))
logger.debug(OUTPUT_SEPARATOR)
logger.debug("DATA")
logger.debug(group)
logger.debug("ROC curve: convex")
logger.debug(roc_convex_hull)
if plot:
from ._curve_plotting_utilities import plot_curve
plot_curve(sensitive_feature_value, 'selection', 'error',
selection_error_curve[sensitive_feature_value])
# Find minimum error point given that at each point the selection rate for each sensitive
# feature value is identical by design.
i_best_DP = error_given_selection.idxmin()
x_best = x_grid[i_best_DP]
# create the solution as interpolation of multiple points with a separate predictor per
# sensitive feature value
predicted_DP_by_sensitive_feature = {}
for sensitive_feature_value in selection_error_curve.keys():
# For DP we already have the predictor directly without complex interpolation.
selection_error_curve_result = selection_error_curve[sensitive_feature_value].transpose()[
i_best_DP]
predicted_DP_by_sensitive_feature[sensitive_feature_value] = \
InterpolatedPredictor(0, 0,
selection_error_curve_result.p0,
selection_error_curve_result.operation0,
selection_error_curve_result.p1,
selection_error_curve_result.operation1)
logger.debug(OUTPUT_SEPARATOR)
logger.debug("From ROC curves")
logger.debug("Best DP: error=%.3f, selection rate=%.3f",
error_given_selection[i_best_DP], x_best)
logger.debug(OUTPUT_SEPARATOR)
if plot:
from ._curve_plotting_utilities import plot_solution_and_show_plot
plot_solution_and_show_plot(
x_best, None, "DP solution", "selection rate", "error")
return predicted_DP_by_sensitive_feature
def _threshold_optimization_equalized_odds(sensitive_features, labels, scores, grid_size=1000,
flip=True, plot=False):
"""Calculate the ROC curve of every sensitive feature value at different thresholds.
Subsequently takes the overlapping region of the ROC curves, and finds the best
solution by selecting the point on the curve with minimal error.
This method assumes that sensitive_features, labels, and scores are non-empty data structures
of equal length, and labels contains only binary labels 0 and 1.
:param sensitive_features: the feature data that determines the groups for which the parity
constraints are enforced
:type sensitive_features: list, numpy.ndarray, pandas.DataFrame, or pandas.Series
:param labels: the labels of the dataset
:type labels: list, numpy.ndarray, pandas.DataFrame, or pandas.Series
:param scores: the scores produced by a predictor's prediction
:type scores: list, numpy.ndarray, pandas.DataFrame, or pandas.Series
:param grid_size: The number of ticks on the grid over which we evaluate the curves.
A large grid_size means that we approximate the actual curve, so it increases the chance
of being very close to the actual best solution.
:type grid_size: int
:param flip: allow flipping to negative weights if it improves accuracy.
:type flip: bool
:param plot: show ROC plot if True
:type plot: bool
:return: the postprocessed predictor as a function taking the sensitive feature value
and the fairness unaware predictor's score as arguments to produce predictions
"""
data_grouped_by_sensitive_feature = _reformat_and_group_data(
sensitive_features, labels, scores)
n = len(labels)
if type(labels) == pd.DataFrame:
n_positive = labels.sum().loc[0]
else:
n_positive = sum(labels)
n_negative = n - n_positive
roc = {}
x_grid = np.linspace(0, 1, grid_size + 1)
y_values = pd.DataFrame()
for sensitive_feature_value, group in data_grouped_by_sensitive_feature:
roc_convex_hull = _get_roc(group, sensitive_feature_value, flip=flip)
roc[sensitive_feature_value] = \
_interpolate_curve(roc_convex_hull, 'x', 'y', 'operation', x_grid)
y_values[sensitive_feature_value] = roc[sensitive_feature_value]['y']
logger.debug(OUTPUT_SEPARATOR)
logger.debug("Processing %s", str(sensitive_feature_value))
logger.debug(OUTPUT_SEPARATOR)
logger.debug("DATA")
logger.debug(group)
logger.debug("ROC curve: convex")
logger.debug(roc_convex_hull)
if plot:
from ._curve_plotting_utilities import plot_curve
plot_curve(sensitive_feature_value, 'x', 'y', roc[sensitive_feature_value])
# Calculate the overlap of the ROC curves by taking the lowest y value
# at every given x.
y_min = np.amin(y_values, axis=1)
# Calculate the error at any given x as the sum of
# a) the proportion of negative labels multiplied by x which represents
# the conditional probability P[Y_hat=1 | Y=0], i.e. the probability
# of a positive prediction given a negative label.
# b) the propotion of positive labels multiplied by 1-y_min, where y_min
# represents the conditional probability P[Y_hat=1 | Y=1], i.e. the
# probability of a correct prediction of a positive label, so 1-y_min
# represents a negative prediction given a positive label.
error_given_x = (n_negative / n) * x_grid + (n_positive / n) * (1 - y_min)
i_best_EO = error_given_x.idxmin()
x_best = x_grid[i_best_EO]
y_best = y_min[i_best_EO]
# create the solution as interpolation of multiple points with a separate predictor
# per sensitive feature
predicted_EO_by_sensitive_feature = {}
for sensitive_feature_value in roc.keys():
roc_result = roc[sensitive_feature_value].transpose()[i_best_EO]
# p_ignore * x_best represent the diagonal of the ROC plot.
if roc_result.y == roc_result.x:
# result is on the diagonal of the ROC plot, i.e. p_ignore is not required
p_ignore = 0
else:
# Calculate p_ignore to change prediction P to y_best
# p_ignore * x_best + (1 - p_ignore) * P
difference_from_best_predictor_for_sensitive_feature = roc_result.y - y_best
vertical_distance_from_diagonal = roc_result.y - roc_result.x
p_ignore = difference_from_best_predictor_for_sensitive_feature / \
vertical_distance_from_diagonal
predicted_EO_by_sensitive_feature[sensitive_feature_value] = \
InterpolatedPredictor(p_ignore, x_best, roc_result.p0, roc_result.operation0,
roc_result.p1, roc_result.operation1)
logger.debug(OUTPUT_SEPARATOR)
logger.debug("From ROC curves")
logger.debug("Best EO: error=%.3f}, FP rate=%.3f}, TP rate=%.3f}",
error_given_x[i_best_EO], x_best, y_best)
logger.debug(OUTPUT_SEPARATOR)
if plot:
from ._curve_plotting_utilities import plot_overlap, plot_solution_and_show_plot
plot_overlap(x_grid, y_min)
plot_solution_and_show_plot(x_best, y_best, 'EO solution', "$P[\\hat{Y}=1|Y=0]$",
"$P[\\hat{Y}=1|Y=1]$")
return predicted_EO_by_sensitive_feature
def _vectorized_prediction(function_dict, sensitive_features, scores):
"""Make predictions for all samples with all provided functions.
However, only use the results from the function that corresponds to the
sensitive feature value of the sample.
This method assumes that sensitive_features and scores are of equal length.
:param function_dict: the functions that apply to various sensitive feature values
:type function_dict: dictionary of functions
:param sensitive_features: the feature data that determines the grouping
:type sensitive_features: list, numpy.ndarray, pandas.DataFrame, or pandas.Series
:param scores: vector of predicted values
:type scores: list, numpy.ndarray, pandas.DataFrame, or pandas.Series
"""
# handle type conversion to ndarray for other types
sensitive_features_vector = _convert_to_ndarray(
sensitive_features, MULTIPLE_DATA_COLUMNS_ERROR_MESSAGE.format("sensitive_features"))
scores_vector = _convert_to_ndarray(scores, SCORES_DATA_TOO_MANY_COLUMNS_ERROR_MESSAGE)
return sum([(sensitive_features_vector == a) * function_dict[a].predict(scores_vector)
for a in function_dict])
def _convert_to_ndarray(data, dataframe_multiple_columns_error_message):
"""Convert the input data from list, pandas.Series, or pandas.DataFrame to numpy.ndarray.
:param data: the data to be converted into a numpy.ndarray
:type data: numpy.ndarray, pandas.Series, pandas.DataFrame, or list
:param dataframe_multiple_columns_error_message: the error message to show in case the
provided data is more than 1-dimensional
:type dataframe_multiple_columns_error_message:
:return: the input data formatted as numpy.ndarray
:rtype: numpy.ndarray
"""
if type(data) == list:
data = np.array(data)
elif type(data) == pd.DataFrame:
if len(data.columns) > 1:
# TODO: extend to multiple columns for additional group data
raise ValueError(dataframe_multiple_columns_error_message)
data = data[data.columns[0]].values
elif type(data) == pd.Series:
data = data.values
return data
def _reformat_and_group_data(sensitive_features, labels, scores, sensitive_feature_names=None):
"""Reformats the data into a new pandas.DataFrame and group by sensitive feature values.
The data are provided as three arguments (`sensitive_features`, `labels`, `scores`) and
the new DataFrame is grouped by sensitive feature values so that subsequently each group
can be handled separately.
:param sensitive_features: the sensitive features based on which the grouping is determined;
currently only a single sensitive feature is supported
:type sensitive_features: pandas.Series, pandas.DataFrame, numpy.ndarray, or list
:param labels: the training labels
:type labels: pandas.Series, pandas.DataFrame, numpy.ndarray, or list
:param scores: the output from the unconstrained predictor used for training the mitigator
:type scores: pandas.Series, pandas.DataFrame, numpy.ndarray, or list
:param sensitive_feature_names: list of names for the sensitive features in case they were
not implicitly provided (e.g. if `sensitive_features` is of type DataFrame); default
None
:type sensitive_feature_names: list of strings
:return: the training data for the mitigator, grouped by sensitive feature value
:rtype: pandas.DataFrameGroupBy
"""
data_dict = {}
# TODO: extend to multiple columns for additional group data
# and name columns after original column names if possible
# or store the original column names
sensitive_feature_name = SENSITIVE_FEATURE_KEY
if sensitive_feature_names is not None:
if sensitive_feature_name in [SCORE_KEY, LABEL_KEY]:
raise ValueError(SENSITIVE_FEATURE_NAME_CONFLICT_DETECTED_ERROR_MESSAGE)
sensitive_feature_name = sensitive_feature_names[0]
_reformat_data_into_dict(sensitive_feature_name, data_dict, sensitive_features)
_reformat_data_into_dict(SCORE_KEY, data_dict, scores)
_reformat_data_into_dict(LABEL_KEY, data_dict, labels)
return pd.DataFrame(data_dict).groupby(sensitive_feature_name)
def _reformat_data_into_dict(key, data_dict, additional_data):
"""Add `additional_data` to `data_dict` with key `key`.
Before `additional_data` is added to `data_dict` it is first
reformatted into a numpy.ndarray or list.
:param key: the key in `data_dict` at which `additional_data` should be stored;
`key` should describe the purpose of `additional_data` in `data_dict`
:type key: str
:param data_dict: the dictionary containing all the relevant data; `additional_data` will be
inserted at the key `key`.
:type data_dict: dict
:param additional_data: the data to be added to `data_dict` at the specified `key`
:type additional_data: numpy.ndarray, pandas.DataFrame, pandas.Series, or list
"""
if type(additional_data) == np.ndarray:
if len(additional_data.shape) > 2 or (len(additional_data.shape) == 2 and
additional_data.shape[1] > 1):
# TODO: extend to multiple columns for additional_group data
raise ValueError(
MULTIPLE_DATA_COLUMNS_ERROR_MESSAGE.format("sensitive_features"))
else:
data_dict[key] = additional_data.reshape(-1)
elif type(additional_data) == pd.DataFrame:
# TODO: extend to multiple columns for additional_data by using column names
for attribute_column in additional_data.columns:
data_dict[key] = additional_data[attribute_column].values
elif type(additional_data) == pd.Series:
data_dict[key] = additional_data.values
elif type(additional_data) == list:
if type(additional_data[0]) == list:
if len(additional_data[0]) > 1:
# TODO: extend to multiple columns for additional_data
raise ValueError(
MULTIPLE_DATA_COLUMNS_ERROR_MESSAGE.format("sensitive_features"))
data_dict[key] = map(lambda a: a[0], additional_data)
else:
data_dict[key] = additional_data
else:
raise TypeError(UNEXPECTED_DATA_TYPE_ERROR_MESSAGE.format(
type(additional_data)))