def test_disparate_impact_remover_np_num(self): fairness_info = { "favorable_labels": [1.0], "protected_attributes": [{"feature": 57, "privileged_groups": [1.0]},], } trainable = DisparateImpactRemover( sensitive_attribute=57 ) >> LogisticRegression(max_iter=1000) train_X = self.creditg_np_num["train_X"] train_y = self.creditg_np_num["train_y"] trained = trainable.fit(train_X, train_y) test_X = self.creditg_np_num["test_X"] test_y = self.creditg_np_num["test_y"] disparate_impact_scorer = lale.lib.aif360.disparate_impact(**fairness_info) impact = disparate_impact_scorer(trained, test_X, test_y) self.assertTrue(0.9 < impact < 1.1, f"impact {impact}")
def test_disparate_impact_remover_pd_cat_no_redact(self): fairness_info = self.creditg_pd_cat["fairness_info"] trainable_remi = DisparateImpactRemover( **fairness_info, redact=False, preparation=self.prep_pd_cat) >> LogisticRegression(max_iter=1000) self._attempt_remi_creditg_pd_cat(fairness_info, trainable_remi, 0.65, 0.75)
def test_disparate_impact_remover_np_num(self): fairness_info = self.creditg_np_num["fairness_info"] trainable_orig = LogisticRegression(max_iter=1000) trainable_remi = DisparateImpactRemover( **fairness_info) >> trainable_orig train_X = self.creditg_np_num["train_X"] train_y = self.creditg_np_num["train_y"] trained_orig = trainable_orig.fit(train_X, train_y) trained_remi = trainable_remi.fit(train_X, train_y) test_X = self.creditg_np_num["test_X"] test_y = self.creditg_np_num["test_y"] disparate_impact_scorer = lale.lib.aif360.disparate_impact( **fairness_info) impact_orig = disparate_impact_scorer(trained_orig, test_X, test_y) self.assertTrue(0.6 < impact_orig < 1.0, f"impact_orig {impact_orig}") impact_remi = disparate_impact_scorer(trained_remi, test_X, test_y) self.assertTrue(0.8 < impact_remi < 1.0, f"impact_remi {impact_remi}")
def test_disparate_impact_remover_pd_cat(self): fairness_info = self.creditg_pd_cat["fairness_info"] trainable_remi = DisparateImpactRemover( **fairness_info, preprocessing=self.prep_pd_cat) >> LogisticRegression( max_iter=1000) self._attempt_remi_creditg_pd_cat(fairness_info, trainable_remi, 0.78, 0.88)
def test_stacking_pre_estimator_mitigation_base_only(self): model = StackingClassifier(estimators=[ ( "dir+dtc", DisparateImpactRemover( **self.fairness_info) >> DecisionTreeClassifier(), ), ("lr", LogisticRegression()), ]) self._attempt_fit_predict(model)
def test_stacking_pre_estimator_mitigation_final_only(self): model = StackingClassifier( estimators=[ ("dtc", DecisionTreeClassifier()), ("lr", LogisticRegression()), ], final_estimator=DisparateImpactRemover(**self.fairness_info) >> DecisionTreeClassifier(), passthrough=True, ) self._attempt_fit_predict(model)
def test_adaboost_pre_estimator_mitigation_base(self): model = AdaBoostClassifier(base_estimator=DisparateImpactRemover( **self.fairness_info) >> DecisionTreeClassifier()) self._attempt_fit_predict(model)
def test_bagging_pre_estimator_mitigation_ensemble(self): model = DisparateImpactRemover( **self.fairness_info) >> BaggingClassifier( base_estimator=DecisionTreeClassifier()) self._attempt_fit_predict(model)
def test_stacking_pre_estimator_mitigation_ensemble(self): model = DisparateImpactRemover( **self.fairness_info) >> StackingClassifier(estimators=[( "dtc", DecisionTreeClassifier()), ("lr", LogisticRegression())]) self._attempt_fit_predict(model)