def __fair_regression_measures(
y_true: ArrayLike, y_pred: ArrayLike, pa_name: str, priv_grp: int = 1
):
""" Returns dict of regression-specific fairness measures
"""
def predmean(_, y_pred, *args):
return np.mean(y_pred.values)
def meanerr(y_true, y_pred, *args):
return np.mean((y_pred - y_true).values)
#
measures = {}
# Ratios
measures["Mean Prediction Ratio"] = aif.ratio(
predmean, y_true, y_pred, prot_attr=pa_name, priv_group=priv_grp
)
measures["MAE Ratio"] = aif.ratio(
mean_absolute_error, y_true, y_pred, prot_attr=pa_name, priv_group=priv_grp
)
# Differences
measures["Mean Prediction Difference"] = aif.difference(
predmean, y_true, y_pred, prot_attr=pa_name, priv_group=priv_grp
)
measures["MAE Difference"] = aif.difference(
mean_absolute_error, y_true, y_pred, prot_attr=pa_name, priv_group=priv_grp
)
return measures
def __fair_classification_measures(
y_true: ArrayLike, y_pred: ArrayLike, pa_name: str, priv_grp: int = 1
):
""" Returns a dict of classification-specific fairness measures
"""
def predmean(_, y_pred, *args):
return np.mean(y_pred.values)
#
measures = {}
measures["Selection Ratio"] = aif.ratio(
predmean, y_true, y_pred, prot_attr=pa_name, priv_group=priv_grp
)
measures["PPV Ratio"] = fcmtrc.ppv_ratio(y_true, y_pred, pa_name, priv_grp)
measures["TPR Ratio"] = fcmtrc.tpr_ratio(y_true, y_pred, pa_name, priv_grp)
measures["FPR Ratio"] = fcmtrc.fpr_ratio(y_true, y_pred, pa_name, priv_grp)
#
measures["Selection Diff"] = aif.difference(
predmean, y_true, y_pred, prot_attr=pa_name, priv_group=priv_grp
)
measures["PPV Diff"] = fcmtrc.ppv_diff(y_true, y_pred, pa_name, priv_grp)
measures["TPR Diff"] = fcmtrc.tpr_diff(y_true, y_pred, pa_name, priv_grp)
measures["FPR Diff"] = fcmtrc.fpr_diff(y_true, y_pred, pa_name, priv_grp)
measures["Balanced Accuracy Difference"] = aif.difference(
balanced_accuracy_score, y_true, y_pred, prot_attr=pa_name, priv_group=priv_grp
)
measures["Balanced Accuracy Ratio"] = aif.ratio(
balanced_accuracy_score, y_true, y_pred, prot_attr=pa_name, priv_group=priv_grp
)
return measures
def __binary_group_fairness_measures(X,
prtc_attr,
y_true,
y_pred,
y_prob=None,
priv_grp=1):
"""[summary]
Args:
X (pandas DataFrame): Sample features
prtc_attr (named array-like): values for the protected attribute
(note: protected attribute may also be present in X)
y_true (pandas DataFrame): Sample targets
y_pred (pandas DataFrame): Sample target predictions
y_prob (pandas DataFrame, optional): Sample target probabilities. Defaults
to None.
Returns:
[type]: [description]
"""
pa_names = prtc_attr.columns.tolist()
gf_vals = {}
gf_key = 'Group Fairness'
gf_vals['Statistical Parity Difference'] = \
aif_mtrc.statistical_parity_difference(y_true, y_pred, prot_attr=pa_names)
gf_vals['Disparate Impact Ratio'] = \
aif_mtrc.disparate_impact_ratio(y_true, y_pred, prot_attr=pa_names)
if not helper.is_tutorial_running() and not len(pa_names) > 1:
gf_vals['Demographic Parity Difference'] = \
fl_mtrc.demographic_parity_difference(y_true, y_pred,
sensitive_features=prtc_attr)
gf_vals['Demographic Parity Ratio'] = \
fl_mtrc.demographic_parity_ratio(y_true, y_pred,
sensitive_features=prtc_attr)
gf_vals['Average Odds Difference'] = \
aif_mtrc.average_odds_difference(y_true, y_pred, prot_attr=pa_names)
gf_vals['Equal Opportunity Difference'] = \
aif_mtrc.equal_opportunity_difference(y_true, y_pred, prot_attr=pa_names)
if not helper.is_tutorial_running() and not len(pa_names) > 1:
gf_vals['Equalized Odds Difference'] = \
fl_mtrc.equalized_odds_difference(y_true, y_pred,
sensitive_features=prtc_attr)
gf_vals['Equalized Odds Ratio'] = \
fl_mtrc.equalized_odds_ratio(y_true, y_pred,
sensitive_features=prtc_attr)
gf_vals['Positive Predictive Parity Difference'] = \
aif_mtrc.difference(sk_metric.precision_score, y_true,
y_pred, prot_attr=pa_names, priv_group=priv_grp)
gf_vals['Balanced Accuracy Difference'] = \
aif_mtrc.difference(sk_metric.balanced_accuracy_score, y_true,
y_pred, prot_attr=pa_names, priv_group=priv_grp)
if y_prob is not None:
gf_vals['AUC Difference'] = \
aif_mtrc.difference(sk_metric.roc_auc_score, y_true, y_prob,
prot_attr=pa_names, priv_group=priv_grp)
return (gf_key, gf_vals)
def update_summary(
summary_dict: Dict[str, Number],
pa_name: str,
y_true: ArrayLike,
y_pred: ArrayLike,
y_prob: ArrayLike,
priv_grp: int,
):
""" Adds replaces measure keys with the names found in the literature
Args:
X (pandas DataFrame): Sample features
pa_name (str):
y_true (pandas DataFrame): Sample targets
y_pred (pandas DataFrame): Sample target predictions
y_prob (pandas DataFrame, optional): Sample target probabilities.
Defaults to None.
priv_grp (int): Specifies which label indicates the privileged
group. Defaults to 1.
"""
name_update = {
"Selection Diff": "Statistical Parity Difference",
"Selection Ratio": "Disparate Impact Ratio",
"PPV Diff": "Positive Predictive Parity Difference",
"PPV Ratio": "Positive Predictive Parity Ratio",
}
drop_keys = ["TPR Ratio", "TPR Diff", "FPR Ratio", "FPR Diff"]
for k in name_update.keys():
val = summary_dict.pop(k)
summary_dict[name_update[k]] = val
for k in drop_keys:
summary_dict.pop(k)
summary_dict["Equal Odds Difference"] = fcmtrc.eq_odds_diff(
y_true, y_pred, pa_name=pa_name
)
summary_dict["Equal Odds Ratio"] = fcmtrc.eq_odds_ratio(
y_true, y_pred, pa_name=pa_name
)
if y_prob is not None:
try:
summary_dict["AUC Difference"] = aif.difference(
pmtrc.roc_auc_score,
y_true,
y_prob,
prot_attr=pa_name,
priv_group=priv_grp,
)
except:
pass
return summary_dict
def fnr_diff(y_true: pd.Series,
y_pred: pd.Series,
pa_name: str,
priv_grp: int = 1):
""" Returns the between-group difference of False Negative Rates
Args:
y_true (pd.Series): true target values
y_pred (pd.Series): predicted target values
prtc_attr (str): name of the protected attribute
priv_grp (int, optional): . Defaults to 1.
Returns:
Number
"""
return difference(false_negative_rate,
y_true,
y_pred,
prot_attr=pa_name,
priv_group=priv_grp)
def ppv_diff(y_true: pd.Series,
y_pred: pd.Series,
pa_name: str,
priv_grp: int = 1):
""" Returns the between-group difference of Positive Predictive Values
Args:
y_true (pd.Series): true target values
y_pred (pd.Series): predicted target values
prtc_attr (str): name of the protected attribute
priv_grp (int, optional): . Defaults to 1.
Returns:
Number
"""
return difference(precision,
y_true,
y_pred,
prot_attr=pa_name,
priv_group=priv_grp)
def score(self, X, y, sample_weight=None):
"""Score the predictions according to the cost constraint specified.
Args:
X (pandas.DataFrame): Probability estimates of the targets as
returned by a ``predict_proba()`` call or equivalent. Note: must
include protected attributes in the index.
y (array-like): Ground-truth (correct) target values.
sample_weight (array-like, optional): Sample weights.
Returns:
float: Absolute value of the difference in cost function for the two
groups (e.g. :func:`~aif360.sklearn.metrics.generalized_fpr` if
``self.cost_constraint`` is 'fpr')
"""
check_is_fitted(self, ['classes_', 'pos_label_'])
pos_idx = np.nonzero(self.classes_ == self.pos_label_)[0][0]
probas_pred = self.predict_proba(X)[:, pos_idx]
return abs(difference(self._weighted_cost, y, probas_pred,
prot_attr=self.prot_attr_, priv_group=self.groups_[1],
pos_label=self.pos_label_, sample_weight=sample_weight))