def calculate_bias_measures(data_orig_train, data_orig_vt, unprivileged_groups, privileged_groups): model = RandomForestClassifier().fit( data_orig_train.features, data_orig_train.labels.ravel(), sample_weight=data_orig_train.instance_weights) dataset = data_orig_vt dataset_pred = dataset.copy() dataset_pred.labels = model.predict(data_orig_vt.features) classified_metric_race = ClassificationMetric( dataset, dataset_pred, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) metric_pred_race = BinaryLabelDatasetMetric( dataset_pred, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) print("Mean difference {}".format(metric_pred_race.mean_difference())) print("Disparate Metric {}".format(metric_pred_race.disparate_impact())) print("Equal Opportunity Difference {}".format( classified_metric_race.equal_opportunity_difference())) print("Average Abs Odds Difference {}".format( classified_metric_race.average_abs_odds_difference())) print("Theil index {}".format(classified_metric_race.theil_index()))
def compute_metrics(dataset_true, dataset_pred, unprivileged_groups, privileged_groups, disp=True): """ Compute the key metrics """ classified_metric_pred = ClassificationMetric( dataset_true, dataset_pred, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) metrics = OrderedDict() metrics["Balanced accuracy"] = 0.5 * ( classified_metric_pred.true_positive_rate() + classified_metric_pred.true_negative_rate()) metrics[ "Statistical parity difference"] = classified_metric_pred.statistical_parity_difference( ) metrics["Disparate impact"] = classified_metric_pred.disparate_impact() metrics[ "Average odds difference"] = classified_metric_pred.average_odds_difference( ) metrics[ "Equal opportunity difference"] = classified_metric_pred.equal_opportunity_difference( ) metrics["Theil index"] = classified_metric_pred.theil_index() if disp: for k in metrics: print("%s = %.4f" % (k, metrics[k])) return metrics
def get_metric_reports(true_dataset,classfied_dataset,privileged_groups,unprivileged_groups): mirror_dataset=classfied_dataset.copy(deepcopy=True) mirror_dataset.labels=copy.deepcopy(true_dataset.labels) metric=ClassificationMetric( dataset=mirror_dataset, classified_dataset=classfied_dataset, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) #Measuring unfairness end report=OrderedDict() report['TPR']=metric.true_positive_rate() report['TNR']=metric.true_negative_rate() report['FPR']=metric.false_positive_rate() report['FNR']=metric.false_negative_rate() report['Balanced_Acc']=0.5*(report['TPR']+report['TNR']) report['Acc']=metric.accuracy() report["Statistical parity difference"]=metric.statistical_parity_difference() report["Disparate impact"]=metric.disparate_impact() report["Equal opportunity difference"]=metric.equal_opportunity_difference() report["Average odds difference"]=metric.average_odds_difference() report["Theil index"]=metric.theil_index() report["United Fairness"]=metric.generalized_entropy_index() return report
def fair_metrics(dataset, pred, pred_is_dataset=False): if pred_is_dataset: dataset_pred = pred else: dataset_pred = dataset.copy() dataset_pred.labels = pred cols = [ 'statistical_parity_difference', 'equal_opportunity_difference', 'average_abs_odds_difference', 'disparate_impact', 'theil_index' ] obj_fairness = [[0, 0, 0, 1, 0]] fair_metrics = pd.DataFrame(data=obj_fairness, index=['objective'], columns=cols) for attr in dataset_pred.protected_attribute_names: idx = dataset_pred.protected_attribute_names.index(attr) privileged_groups = [{ attr: dataset_pred.privileged_protected_attributes[idx][0] }] unprivileged_groups = [{ attr: dataset_pred.unprivileged_protected_attributes[idx][0] }] classified_metric = ClassificationMetric( dataset, dataset_pred, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) metric_pred = BinaryLabelDatasetMetric( dataset_pred, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) acc = classified_metric.accuracy() row = pd.DataFrame([[ metric_pred.mean_difference(), classified_metric.equal_opportunity_difference(), classified_metric.average_abs_odds_difference(), metric_pred.disparate_impact(), classified_metric.theil_index() ]], columns=cols, index=[attr]) fair_metrics = fair_metrics.append(row) fair_metrics = fair_metrics.replace([-np.inf, np.inf], 2) return fair_metrics
def test(dataset, model, x_test, thresh_arr, unprivileged_groups, privileged_groups): bld = BinaryLabelDataset(df=dataset, label_names=['labels'], protected_attribute_names=['age']) if np.isin(k, model_AIF): y_val_pred_prob = model.predict_proba(bld) else: y_val_pred_prob, A_val_pred_prob = model.predict_proba(x_test) metric_arrs = np.empty([0, 8]) for thresh in thresh_arr: if np.isin(k, model_AIF): y_val_pred = (y_val_pred_prob > thresh).astype(np.float64) else: y_val_pred = (y_val_pred_prob.numpy() > thresh).astype(np.float64) metric_arrs = np.append(metric_arrs, roc_auc_score(y_test, y_val_pred_prob)) if np.isin(k, model_AIF): metric_arrs = np.append(metric_arrs, 0) else: metric_arrs = np.append(metric_arrs, roc_auc_score(A_test, A_val_pred_prob)) dataset_pred = dataset.copy() dataset_pred.labels = y_val_pred bld2 = BinaryLabelDataset(df=dataset_pred, label_names=['labels'], protected_attribute_names=['age']) metric = ClassificationMetric(bld, bld2, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) metric_arrs = np.append( metric_arrs, ((metric.true_positive_rate() + metric.true_negative_rate()) / 2)) metric_arrs = np.append(metric_arrs, metric.average_odds_difference()) metric_arrs = np.append(metric_arrs, metric.disparate_impact()) metric_arrs = np.append(metric_arrs, metric.statistical_parity_difference()) metric_arrs = np.append(metric_arrs, metric.equal_opportunity_difference()) metric_arrs = np.append(metric_arrs, metric.theil_index()) return metric_arrs
def nondebiased_classifier(train, test, privileged_groups, unprivileged_groups): sess = tf.Session() NN_model = AdversarialDebiasing(privileged_groups, unprivileged_groups, scope_name='nondebiased_classifier', debias=False, sess=sess) NN_model.fit(train) # predict outcome using the test set pred_NNmodel = NN_model.predict(test) sess.close() tf.reset_default_graph() # calculate accuracy accuracy = accuracy_score(y_true=test.labels, y_pred=pred_NNmodel.labels) # calculate fairness metrics metric_test = BinaryLabelDatasetMetric( pred_NNmodel, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) acc_test = ClassificationMetric(test, pred_NNmodel, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) equal_opportunity_difference = equal_opp_diff(test, pred_NNmodel, 'sex', privileged=1, unprivileged=0, favourable=1, unfavourable=0) average_odds_difference = avg_odds_diff(test, pred_NNmodel, 'sex', privileged=1, unprivileged=0, favourable=1, unfavourable=0) metrics = [ metric_test.mean_difference(), acc_test.disparate_impact(), equal_opportunity_difference, average_odds_difference, acc_test.theil_index() ] return pred_NNmodel, accuracy, metrics
def test_theil_index(): data = np.array([[0, 1], [0, 0], [1, 0], [1, 1], [1, 0], [1, 0], [2, 1], [2, 0], [2, 1], [2, 1]]) pred = data.copy() pred[[3, 9], -1] = 0 pred[[4, 5], -1] = 1 df = pd.DataFrame(data, columns=['feat', 'label']) df2 = pd.DataFrame(pred, columns=['feat', 'label']) bld = BinaryLabelDataset(df=df, label_names=['label'], protected_attribute_names=['feat']) bld2 = BinaryLabelDataset(df=df2, label_names=['label'], protected_attribute_names=['feat']) cm = ClassificationMetric(bld, bld2) assert cm.theil_index() == 4 * np.log(2) / 10
def prejudice(train, test, unprivileged_groups, privileged_groups): prejudice_model = PrejudiceRemover(eta=100, sensitive_attr='sex') prejudice_model.fit(train) # predict outcome using the test set pred_prejudice = prejudice_model.predict(test) # calculate accuracy accuracy = accuracy_score(y_true=test.labels, y_pred=pred_prejudice.labels) # calculate fairness metrics metric_test = BinaryLabelDatasetMetric( pred_prejudice, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) acc_test = ClassificationMetric(test, pred_prejudice, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) equal_opportunity_difference = equal_opp_diff(test, pred_prejudice, 'sex', privileged=1, unprivileged=0, favourable=1, unfavourable=0) average_odds_difference = avg_odds_diff(test, pred_prejudice, 'sex', privileged=1, unprivileged=0, favourable=1, unfavourable=0) if acc_test.disparate_impact() == math.inf: disparate_impact = 5.0 else: disparate_impact = acc_test.disparate_impact() metrics = [ metric_test.mean_difference(), disparate_impact, equal_opportunity_difference, average_odds_difference, acc_test.theil_index() ] return pred_prejudice, accuracy, metrics
def ensemble(test, pred_adversarial, pred_prejudice, pred_nondebiased, unprivileged_groups, privileged_groups): pred_labels = [] for i in range(0, len(test.features)): arr = mode([ pred_adversarial.labels[i], pred_prejudice.labels[i], pred_nondebiased.labels[i] ]) pred_labels.append(arr[0][0]) pred_ensemble = test.copy() pred_ensemble.labels = np.array(pred_labels) accuracy = accuracy_score(y_true=test.labels, y_pred=pred_ensemble.labels) metric_test = BinaryLabelDatasetMetric( pred_ensemble, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) acc_test = ClassificationMetric(test, pred_ensemble, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) equal_opportunity_difference = equal_opp_diff(test, pred_ensemble, 'sex', privileged=1, unprivileged=0, favourable=1, unfavourable=0) average_odds_difference = avg_odds_diff(test, pred_ensemble, 'sex', privileged=1, unprivileged=0, favourable=1, unfavourable=0) metrics = [ metric_test.mean_difference(), acc_test.disparate_impact(), equal_opportunity_difference, average_odds_difference, acc_test.theil_index() ] return accuracy, metrics
def compute_aif_metrics(dataset_true, dataset_pred, unprivileged_groups, privileged_groups,\ ret_eval_dict=True): metrics_cls = ClassificationMetric(dataset_true, dataset_pred, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) metrics_dict = {} metrics_dict["BA"] = 0.5*(metrics_cls.true_positive_rate()+ metrics_cls.true_negative_rate()) metrics_dict["SPD"] = metrics_cls.statistical_parity_difference() metrics_dict["DI"] = metrics_cls.disparate_impact() metrics_dict["AOD"] = metrics_cls.average_odds_difference() metrics_dict["EOD"] = metrics_cls.equal_opportunity_difference() metrics_dict["DFBA"] = metrics_cls.differential_fairness_bias_amplification() metrics_dict["TI"] = metrics_cls.theil_index() if ret_eval_dict: return metrics_dict, metrics_cls else: return metrics_cls
"#### Plain model - without debiasing - classification metrics on test set" ) # print("Test set: Classification accuracy = %f" % classified_metric_nodebiasing_test.accuracy()) # print("Test set: Balanced classification accuracy = %f" % bal_acc_nodebiasing_test) # print("Test set: Statistical parity difference = %f" % classified_metric_nodebiasing_test.statistical_parity_difference()) # print("Test set: Disparate impact = %f" % classified_metric_nodebiasing_test.disparate_impact()) # print("Test set: Equal opportunity difference = %f" % classified_metric_nodebiasing_test.equal_opportunity_difference()) # print("Test set: Average odds difference = %f" % classified_metric_nodebiasing_test.average_odds_difference()) # print("Test set: Theil index = %f" % classified_metric_nodebiasing_test.theil_index()) # print("Test set: False negative rate difference = %f" % classified_metric_nodebiasing_test.false_negative_rate_difference()) metrics = { "Classification accuracy": classified_metric_nodebiasing_test.accuracy(), "Balanced classification accuracy": bal_acc_nodebiasing_test, "Statistical parity difference": classified_metric_nodebiasing_test.statistical_parity_difference(), "Disparate impact": classified_metric_nodebiasing_test.disparate_impact(), "Equal opportunity difference": classified_metric_nodebiasing_test.equal_opportunity_difference(), "Average odds difference": classified_metric_nodebiasing_test.average_odds_difference(), "Theil index": classified_metric_nodebiasing_test.theil_index(), "False negative rate difference": classified_metric_nodebiasing_test.false_negative_rate_difference() } print("metrics: ", metrics)
def get_fair_metrics(dataset, pred, pred_is_dataset=False): """ Measure fairness metrics. Parameters: dataset (pandas dataframe): Dataset pred (array): Model predictions pred_is_dataset, optional (bool): True if prediction is already part of the dataset, column name 'labels'. Returns: fair_metrics: Fairness metrics. """ if pred_is_dataset: dataset_pred = pred else: dataset_pred = dataset.copy() dataset_pred.labels = pred cols = [ 'statistical_parity_difference', 'equal_opportunity_difference', 'average_abs_odds_difference', 'disparate_impact', 'theil_index' ] obj_fairness = [[0, 0, 0, 1, 0]] fair_metrics = pd.DataFrame(data=obj_fairness, index=['objective'], columns=cols) for attr in dataset_pred.protected_attribute_names: idx = dataset_pred.protected_attribute_names.index(attr) privileged_groups = [{ attr: dataset_pred.privileged_protected_attributes[idx][0] }] unprivileged_groups = [{ attr: dataset_pred.unprivileged_protected_attributes[idx][0] }] classified_metric = ClassificationMetric( dataset, dataset_pred, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) metric_pred = BinaryLabelDatasetMetric( dataset_pred, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) acc = classified_metric.accuracy() row = pd.DataFrame([[ metric_pred.mean_difference(), classified_metric.equal_opportunity_difference(), classified_metric.average_abs_odds_difference(), metric_pred.disparate_impact(), classified_metric.theil_index() ]], columns=cols, index=[attr]) fair_metrics = fair_metrics.append(row) fair_metrics = fair_metrics.replace([-np.inf, np.inf], 2) return fair_metrics
def fairness_check(s3_url, bucket_name, s3_username, s3_password, training_id): cos = boto3.resource("s3", endpoint_url=s3_url, aws_access_key_id=s3_username, aws_secret_access_key=s3_password) y_test_out = 'y_test.out' p_test_out = 'p_test.out' y_pred_out = 'y_pred.out' get_s3_item(cos, bucket_name, training_id + '/' + y_test_out, y_test_out) get_s3_item(cos, bucket_name, training_id + '/' + p_test_out, p_test_out) get_s3_item(cos, bucket_name, training_id + '/' + y_pred_out, y_pred_out) """Need to generalize the protected features""" unprivileged_groups = [{'race': 4.0}] privileged_groups = [{'race': 0.0}] favorable_label = 0.0 unfavorable_label = 1.0 """Load the necessary labels and protected features for fairness check""" y_test = np.loadtxt(y_test_out) p_test = np.loadtxt(p_test_out) y_pred = np.loadtxt(y_pred_out) """Calculate the fairness metrics""" original_test_dataset = dataset_wrapper(outcome=y_test, protected=p_test, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups, favorable_label=favorable_label, unfavorable_label=unfavorable_label) plain_predictions_test_dataset = dataset_wrapper(outcome=y_pred, protected=p_test, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups, favorable_label=favorable_label, unfavorable_label=unfavorable_label) classified_metric_nodebiasing_test = ClassificationMetric(original_test_dataset, plain_predictions_test_dataset, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) TPR = classified_metric_nodebiasing_test.true_positive_rate() TNR = classified_metric_nodebiasing_test.true_negative_rate() bal_acc_nodebiasing_test = 0.5*(TPR+TNR) print("#### Plain model - without debiasing - classification metrics on test set") metrics = { "Classification accuracy": classified_metric_nodebiasing_test.accuracy(), "Balanced classification accuracy": bal_acc_nodebiasing_test, "Statistical parity difference": classified_metric_nodebiasing_test.statistical_parity_difference(), "Disparate impact": classified_metric_nodebiasing_test.disparate_impact(), "Equal opportunity difference": classified_metric_nodebiasing_test.equal_opportunity_difference(), "Average odds difference": classified_metric_nodebiasing_test.average_odds_difference(), "Theil index": classified_metric_nodebiasing_test.theil_index(), "False negative rate difference": classified_metric_nodebiasing_test.false_negative_rate_difference() } print("metrics: ", metrics) return metrics
def comb_algorithm(l, m, n, dataset_original1, privileged_groups1, unprivileged_groups1, optim_options1): dataset_original2 = copy.deepcopy(dataset_original1) privileged_groups2 = copy.deepcopy(privileged_groups1) unprivileged_groups2 = copy.deepcopy(unprivileged_groups1) optim_options2 = copy.deepcopy(optim_options1) print(l, m, n) dataset_original_train, dataset_original_vt = dataset_original2.split( [0.7], shuffle=True) dataset_original_valid, dataset_original_test = dataset_original_vt.split( [0.5], shuffle=True) dataset_original_test.labels = dataset_original_test.labels print('=======================') #print(dataset_original_test.labels) dataset_orig_train = copy.deepcopy(dataset_original_train) dataset_orig_valid = copy.deepcopy(dataset_original_valid) dataset_orig_test = copy.deepcopy(dataset_original_test) if l == 0: dataset_transfer_train = copy.deepcopy(dataset_original_train) dataset_transfer_valid = copy.deepcopy(dataset_original_valid) dataset_transfer_test = copy.deepcopy(dataset_original_test) #dataset_transf_train, dataset_transf_valid, dataset_transf_test = dataset_orig_train, dataset_orig_valid, dataset_orig_test else: pre_used = preAlgorithm[l - 1] dataset_transfer_train, dataset_transfer_valid, dataset_transfer_test = Pre( pre_used, dataset_orig_train, dataset_orig_valid, dataset_orig_test, privileged_groups2, unprivileged_groups2, optim_options2) dataset_transf_train = copy.deepcopy(dataset_transfer_train) dataset_transf_valid = copy.deepcopy(dataset_transfer_valid) dataset_transf_test = copy.deepcopy(dataset_transfer_test) if m == 0: dataset_transfer_valid_pred, dataset_transfer_test_pred = plain_model( dataset_transf_train, dataset_transf_valid, dataset_transf_test, privileged_groups2, unprivileged_groups2) else: in_used = inAlgorithm[m - 1] if in_used == "adversarial_debiasing": dataset_transfer_valid_pred, dataset_transfer_test_pred = adversarial_debiasing( dataset_transf_train, dataset_transf_valid, dataset_transf_test, privileged_groups2, unprivileged_groups2) elif in_used == "art_classifier": dataset_transfer_valid_pred, dataset_transfer_test_pred = art_classifier( dataset_transf_train, dataset_transf_valid, dataset_transf_test, privileged_groups2, unprivileged_groups2) elif in_used == "prejudice_remover": for key, value in privileged_groups2[0].items(): sens_attr = key dataset_transfer_valid_pred, dataset_transfer_test_pred = prejudice_remover( dataset_transf_train, dataset_transf_valid, dataset_transf_test, privileged_groups2, unprivileged_groups2, sens_attr) dataset_transf_valid_pred = copy.deepcopy(dataset_transfer_valid_pred) dataset_transf_test_pred = copy.deepcopy(dataset_transfer_test_pred) if n == 0: dataset_transf_test_pred_transf = copy.deepcopy( dataset_transfer_test_pred) else: post_used = postAlgorithm[n - 1] if post_used == "calibrated_eqodds": cpp = CalibratedEqOddsPostprocessing( privileged_groups=privileged_groups2, unprivileged_groups=unprivileged_groups2, cost_constraint=cost_constraint) cpp = cpp.fit(dataset_transfer_valid, dataset_transf_valid_pred) dataset_transf_test_pred_transf = cpp.predict( dataset_transf_test_pred) elif post_used == "eqodds": EO = EqOddsPostprocessing(unprivileged_groups=unprivileged_groups2, privileged_groups=privileged_groups2) EO = EO.fit(dataset_transfer_valid, dataset_transf_valid_pred) dataset_transf_test_pred_transf = EO.predict( dataset_transf_test_pred) elif post_used == "reject_option": #dataset_transf_test_pred_transf = reject_option(dataset_transf_valid, dataset_transf_valid_pred, dataset_transf_test, dataset_transf_test_pred, privileged_groups2, unprivileged_groups2) ROC = RejectOptionClassification( unprivileged_groups=unprivileged_groups2, privileged_groups=privileged_groups2) ROC = ROC.fit(dataset_transfer_valid, dataset_transf_valid_pred) dataset_transf_test_pred_transf = ROC.predict( dataset_transf_test_pred) #print('=======================') org_labels = dataset_orig_test.labels print(dataset_orig_test.labels) #print(dataset_transf_test.labels) #print('=======================') pred_labels = dataset_transf_test_pred.labels print(dataset_transf_test_pred.labels) true_pred = org_labels == pred_labels print("acc after in: ", float(np.sum(true_pred)) / pred_labels.shape[1]) #print('=======================') #print(dataset_transf_test_pred_transf.labels) #print(dataset_transf_test_pred_transf.labels.shape) metric = ClassificationMetric(dataset_transfer_test, dataset_transf_test_pred_transf, unprivileged_groups=unprivileged_groups2, privileged_groups=privileged_groups2) metrics = OrderedDict() metrics["Classification accuracy"] = metric.accuracy() TPR = metric.true_positive_rate() TNR = metric.true_negative_rate() bal_acc_nodebiasing_test = 0.5 * (TPR + TNR) metrics["Balanced classification accuracy"] = bal_acc_nodebiasing_test metrics[ "Statistical parity difference"] = metric.statistical_parity_difference( ) metrics["Disparate impact"] = metric.disparate_impact() metrics[ "Equal opportunity difference"] = metric.equal_opportunity_difference( ) metrics["Average odds difference"] = metric.average_odds_difference() metrics["Theil index"] = metric.theil_index() metrics["United Fairness"] = metric.generalized_entropy_index() feature = [] feature_str = "[" for m in metrics: data = round(metrics[m], 4) feature.append(data) feature_str = feature_str + str(data) + " " feature_str = feature_str + "]" return feature, feature_str
def fairness_check(label_dir, model_dir): """Need to generalize the protected features""" # races_to_consider = [0,4] unprivileged_groups = [{'race': 4.0}] privileged_groups = [{'race': 0.0}] favorable_label = 0.0 unfavorable_label = 1.0 """Load the necessary labels and protected features for fairness check""" # y_train = np.loadtxt(label_dir + '/y_train.out') # p_train = np.loadtxt(label_dir + '/p_train.out') y_test = np.loadtxt(label_dir + '/y_test.out') p_test = np.loadtxt(label_dir + '/p_test.out') y_pred = np.loadtxt(label_dir + '/y_pred.out') """Calculate the fairness metrics""" # original_traning_dataset = dataset_wrapper(outcome=y_train, protected=p_train, # unprivileged_groups=unprivileged_groups, # privileged_groups=privileged_groups, # favorable_label=favorable_label, # unfavorable_label=unfavorable_label) original_test_dataset = dataset_wrapper(outcome=y_test, protected=p_test, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups, favorable_label=favorable_label, unfavorable_label=unfavorable_label) plain_predictions_test_dataset = dataset_wrapper(outcome=y_pred, protected=p_test, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups, favorable_label=favorable_label, unfavorable_label=unfavorable_label) classified_metric_nodebiasing_test = ClassificationMetric(original_test_dataset, plain_predictions_test_dataset, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) TPR = classified_metric_nodebiasing_test.true_positive_rate() TNR = classified_metric_nodebiasing_test.true_negative_rate() bal_acc_nodebiasing_test = 0.5*(TPR+TNR) print("#### Plain model - without debiasing - classification metrics on test set") # print("Test set: Classification accuracy = %f" % classified_metric_nodebiasing_test.accuracy()) # print("Test set: Balanced classification accuracy = %f" % bal_acc_nodebiasing_test) # print("Test set: Statistical parity difference = %f" % classified_metric_nodebiasing_test.statistical_parity_difference()) # print("Test set: Disparate impact = %f" % classified_metric_nodebiasing_test.disparate_impact()) # print("Test set: Equal opportunity difference = %f" % classified_metric_nodebiasing_test.equal_opportunity_difference()) # print("Test set: Average odds difference = %f" % classified_metric_nodebiasing_test.average_odds_difference()) # print("Test set: Theil index = %f" % classified_metric_nodebiasing_test.theil_index()) # print("Test set: False negative rate difference = %f" % classified_metric_nodebiasing_test.false_negative_rate_difference()) metrics = { "Classification accuracy": classified_metric_nodebiasing_test.accuracy(), "Balanced classification accuracy": bal_acc_nodebiasing_test, "Statistical parity difference": classified_metric_nodebiasing_test.statistical_parity_difference(), "Disparate impact": classified_metric_nodebiasing_test.disparate_impact(), "Equal opportunity difference": classified_metric_nodebiasing_test.equal_opportunity_difference(), "Average odds difference": classified_metric_nodebiasing_test.average_odds_difference(), "Theil index": classified_metric_nodebiasing_test.theil_index(), "False negative rate difference": classified_metric_nodebiasing_test.false_negative_rate_difference() } return {"metrics": metrics}
ClassificationMetric( dataset_ground_truth, dataset_classifier, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) TPR = classificaltion_metric.true_positive_rate() TNR = classificaltion_metric.true_negative_rate() bal_acc_nodebiasing_test = 0.5 * (TPR + TNR) metrics = { "classification_accuracy": classificaltion_metric.accuracy(), "balanced_classification_accuracy": bal_acc_nodebiasing_test, "statistical_parity_difference": classificaltion_metric.statistical_parity_difference(), "disparate_impact": classificaltion_metric.disparate_impact(), "equal_opportunity_difference": classificaltion_metric.equal_opportunity_difference(), "average_odds_difference": classificaltion_metric.average_odds_difference(), "theil_index": classificaltion_metric.theil_index(), "false_negative_rate_difference": classificaltion_metric.false_negative_rate_difference() } sys.stdout.write(json.dumps(metrics))
def fairness_check(object_storage_url, object_storage_username, object_storage_password, data_bucket_name, result_bucket_name, model_id, feature_testset_path='processed_data/X_test.npy', label_testset_path='processed_data/y_test.npy', protected_label_testset_path='processed_data/p_test.npy', model_class_file='model.py', model_class_name='model', favorable_label=0.0, unfavorable_label=1.0, privileged_groups=[{ 'race': 0.0 }], unprivileged_groups=[{ 'race': 4.0 }]): url = re.compile(r"https?://") cos = Minio(url.sub('', object_storage_url), access_key=object_storage_username, secret_key=object_storage_password, secure=False) # Local Minio server won't have HTTPS dataset_filenamex = "X_test.npy" dataset_filenamey = "y_test.npy" dataset_filenamep = "p_test.npy" weights_filename = "model.pt" model_files = model_id + '/_submitted_code/model.zip' cos.fget_object(data_bucket_name, feature_testset_path, dataset_filenamex) cos.fget_object(data_bucket_name, label_testset_path, dataset_filenamey) cos.fget_object(data_bucket_name, protected_label_testset_path, dataset_filenamep) cos.fget_object(result_bucket_name, model_id + '/' + weights_filename, weights_filename) cos.fget_object(result_bucket_name, model_files, 'model.zip') # Load PyTorch model definition from the source code. zip_ref = zipfile.ZipFile('model.zip', 'r') zip_ref.extractall('model_files') zip_ref.close() modulename = 'model_files.' + model_class_file.split('.')[0].replace( '-', '_') ''' We required users to define where the model class is located or follow some naming convention we have provided. ''' model_class = getattr(importlib.import_module(modulename), model_class_name) # load & compile model device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') model = model_class().to(device) model.load_state_dict(torch.load(weights_filename, map_location=device)) """Load the necessary labels and protected features for fairness check""" x_test = np.load(dataset_filenamex) y_test = np.load(dataset_filenamey) p_test = np.load(dataset_filenamep) _, y_pred = evaluate(model, x_test, y_test) """Calculate the fairness metrics""" original_test_dataset = dataset_wrapper( outcome=y_test, protected=p_test, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups, favorable_label=favorable_label, unfavorable_label=unfavorable_label) plain_predictions_test_dataset = dataset_wrapper( outcome=y_pred, protected=p_test, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups, favorable_label=favorable_label, unfavorable_label=unfavorable_label) classified_metric_nodebiasing_test = ClassificationMetric( original_test_dataset, plain_predictions_test_dataset, unprivileged_groups=unprivileged_groups, privileged_groups=privileged_groups) TPR = classified_metric_nodebiasing_test.true_positive_rate() TNR = classified_metric_nodebiasing_test.true_negative_rate() bal_acc_nodebiasing_test = 0.5 * (TPR + TNR) print( "#### Plain model - without debiasing - classification metrics on test set" ) metrics = { "Classification accuracy": classified_metric_nodebiasing_test.accuracy(), "Balanced classification accuracy": bal_acc_nodebiasing_test, "Statistical parity difference": classified_metric_nodebiasing_test.statistical_parity_difference(), "Disparate impact": classified_metric_nodebiasing_test.disparate_impact(), "Equal opportunity difference": classified_metric_nodebiasing_test.equal_opportunity_difference(), "Average odds difference": classified_metric_nodebiasing_test.average_odds_difference(), "Theil index": classified_metric_nodebiasing_test.theil_index(), "False negative rate difference": classified_metric_nodebiasing_test.false_negative_rate_difference() } print("metrics: ", metrics) return metrics
def comb_algorithm(l, m, n, dataset_original1, privileged_groups1, unprivileged_groups1, optim_options1): dataset_original2 = copy.deepcopy(dataset_original1) privileged_groups2 = copy.deepcopy(privileged_groups1) unprivileged_groups2 = copy.deepcopy(unprivileged_groups1) optim_options2 = copy.deepcopy(optim_options1) print(l, m, n) dataset_orig_train, dataset_orig_vt = dataset_original2.split([0.7], shuffle=True) dataset_orig_valid, dataset_orig_test = dataset_orig_vt.split([0.5], shuffle=True) if l == 0: dataset_transf_train, dataset_transf_valid, dataset_transf_test = dataset_orig_train, dataset_orig_valid, dataset_orig_test else: pre_used = preAlgorithm[l - 1] dataset_transf_train, dataset_transf_valid, dataset_transf_test = Pre( pre_used, dataset_orig_train, dataset_orig_valid, dataset_orig_test, privileged_groups2, unprivileged_groups2, optim_options2) #assert (l,m,n)!=(2,0,0) #assert not np.all(dataset_transf_train.labels.flatten()==1.0) if m == 0: dataset_transf_valid_pred, dataset_transf_test_pred = train( dataset_transf_train, dataset_transf_valid, dataset_transf_test, privileged_groups2, unprivileged_groups2) else: in_used = inAlgorithm[m - 1] if in_used == "adversarial_debiasing": dataset_transf_valid_pred, dataset_transf_test_pred = adversarial_debiasing( dataset_transf_train, dataset_transf_valid, dataset_transf_test, privileged_groups2, unprivileged_groups2) elif in_used == "art_classifier": dataset_transf_valid_pred, dataset_transf_test_pred = art_classifier( dataset_transf_train, dataset_transf_valid, dataset_transf_test, privileged_groups2, unprivileged_groups2) elif in_used == "prejudice_remover": for key, value in privileged_groups2[0].items(): sens_attr = key dataset_transf_valid_pred, dataset_transf_test_pred = prejudice_remover( dataset_transf_train, dataset_transf_valid, dataset_transf_test, privileged_groups2, unprivileged_groups2, sens_attr) if n == 0: dataset_transf_test_pred_transf = dataset_transf_test_pred else: post_used = postAlgorithm[n - 1] if post_used == "calibrated_eqodds": cpp = CalibratedEqOddsPostprocessing( privileged_groups=privileged_groups2, unprivileged_groups=unprivileged_groups2, cost_constraint=cost_constraint, seed=1) cpp = cpp.fit(dataset_transf_valid, dataset_transf_valid_pred) dataset_transf_test_pred_transf = cpp.predict( dataset_transf_test_pred) elif post_used == "eqodds": EO = EqOddsPostprocessing(unprivileged_groups=unprivileged_groups2, privileged_groups=privileged_groups2, seed=1) EO = EO.fit(dataset_transf_valid, dataset_transf_valid_pred) dataset_transf_test_pred_transf = EO.predict( dataset_transf_test_pred) elif post_used == "reject_option": ROC = RejectOptionClassification( unprivileged_groups=unprivileged_groups2, privileged_groups=privileged_groups2, low_class_thresh=0.01, high_class_thresh=0.99, num_class_thresh=100, num_ROC_margin=50, metric_name=allowed_metrics[0], metric_ub=metric_ub, metric_lb=metric_lb) ROC = ROC.fit(dataset_transf_valid, dataset_transf_valid_pred) dataset_transf_test_pred_transf = ROC.predict( dataset_transf_test_pred) metric = ClassificationMetric(dataset_transf_test, dataset_transf_test_pred_transf, unprivileged_groups=unprivileged_groups2, privileged_groups=privileged_groups2) metrics = OrderedDict() metrics["Classification accuracy"] = metric.accuracy() TPR = metric.true_positive_rate() TNR = metric.true_negative_rate() bal_acc_nodebiasing_test = 0.5 * (TPR + TNR) metrics["Balanced classification accuracy"] = bal_acc_nodebiasing_test metrics[ "Statistical parity difference"] = metric.statistical_parity_difference( ) metrics["Disparate impact"] = metric.disparate_impact() metrics[ "Equal opportunity difference"] = metric.equal_opportunity_difference( ) metrics["Average odds difference"] = metric.average_odds_difference() metrics["Theil index"] = metric.theil_index() metrics["United Fairness"] = metric.generalized_entropy_index() # print(metrics) feature = "[" for m in metrics: feature = feature + " " + str(round(metrics[m], 4)) feature = feature + "]" return feature