Auditing Bias of Machine Learning Algorithms

In this tutorial, we audit a machine learning model in terms of its fairness property. Our focus is on group fairness, a popular paradigm of fairness notions that tries to establish equity in prediction across different demographic groups in the data based on sensitive features: race, gender, etc. By the end of this notebook, the audience will be demonstrated a formal framework for quantifying bias/unfairness in machine learning and a framework to explain bias in terms of feature combinations in the data.

• Formal Quantification of Bias: We deploy a recently developed formal fairness quantifier Justicia to quantify the bias of a machine learning classifier that relies on the probability distribution of features to measure the bias of the classifier. (https://github.com/meelgroup/justicia)
• Explaining Sources of Bias: We demonstrate a fairness explainer, called FairXplainer to decompose the bias of the classifier among unfairness contributions of different combinations of features. (https://github.com/ReAILe/bias-explainer)
• Fairness Enhancing Algorithms We improve unfairness by applying a pre-processing reweighing-based fairness algorithm from AIF360, a comprehensive library of fairness algorithms (https://github.com/ReAILe/bias-explainer)
• For experiments, we consider recently introduced US census datasets incorporated in a python package folktables. The details of the dataset is here: https://github.com/socialfoundations/folktables

# Necessary libraries
from folktables import ACSDataSource, ACSEmployment, generate_categories
import math
import numpy as np
import matplotlib.pyplot as plt
from sklearn.linear_model import LogisticRegression
from sklearn.model_selection import train_test_split
from sklearn.pipeline import make_pipeline
from sklearn.preprocessing import StandardScaler, MinMaxScaler
from justicia.metrics import Metric
import aif360
from sklearn import svm
from aif360.algorithms.preprocessing import DisparateImpactRemover
from aif360.algorithms.preprocessing.reweighing import Reweighing
from fairxplainer.fair_explainer import FairXplainer
from fairxplainer.fair_explainer import plot as fif_plot
from fairxplainer import utils
from sklearn.metrics import classification_report



fontsize = 18
labelsize = 16



good_name_features = {
        'AGEP' : 'Age',
        'SCHL' : 'Education',
        'MAR' : 'Marital_status',
        'RELP' : 'Relationship',
        'DIS' : 'Disability',
        'ESP' : "Parent's_employment",
        'CIT' : 'Citizenship',
        'MIG' : 'Mobility_status',
        'MIL' : 'Military_service',
        'ANC' : 'Ancestry',
        'NATIVITY' : 'Nativity',
        'DEAR' : 'Hearing_difficulty',
        'DEYE' : 'Vision_difficulty',
        'DREM' : 'Cognitive_difficulty',
        'SEX' : 'Sex',
        'RAC1P' : 'Race'
}

/Users/bishwamittraghosh/mambaforge-pypy3/envs/fairness_tutorial/lib/python3.7/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
  from .autonotebook import tqdm as notebook_tqdm
WARNING:root:No module named 'tempeh': LawSchoolGPADataset will be unavailable. To install, run:
pip install 'aif360[LawSchoolGPA]'
WARNING:root:No module named 'tensorflow': AdversarialDebiasing will be unavailable. To install, run:
pip install 'aif360[AdversarialDebiasing]'
WARNING:root:No module named 'tensorflow': AdversarialDebiasing will be unavailable. To install, run:
pip install 'aif360[AdversarialDebiasing]'
WARNING:root:No module named 'fairlearn': ExponentiatedGradientReduction will be unavailable. To install, run:
pip install 'aif360[Reductions]'
WARNING:root:No module named 'fairlearn': GridSearchReduction will be unavailable. To install, run:
pip install 'aif360[Reductions]'
WARNING:root:No module named 'fairlearn': GridSearchReduction will be unavailable. To install, run:
pip install 'aif360[Reductions]'

Dataset

data_source = ACSDataSource(survey_year='2018', horizon='1-Year', survey='person')
definition_df = data_source.get_definitions(download=True)
acs_data = data_source.get_data(states=["AL"], download=True)
features_unscaled, label, group = ACSEmployment.df_to_pandas(acs_data)
categories_ = generate_categories(features=features_unscaled, definition_df=definition_df)
categories = {}
for key in categories_.keys():    
    categories[good_name_features[key]] = categories_[key]
features_unscaled = features_unscaled.rename(columns=good_name_features)
features_unscaled.shape

(47777, 16)

Preprocessing

features_unscaled.drop(['Citizenship', 'Ancestry', 'Nativity', "Parent's_employment"], axis=1, inplace=True) # dropping a few collumns
flag = (20 <= features_unscaled['Age']) & (features_unscaled['Age'] <= 60) # flagging the rows that have age between 20 and 60, as individuals outside this range are not in the labor force
label = label[flag]
features_unscaled = features_unscaled[flag]

features_unscaled.shape, label.shape

((23793, 12), (23793, 1))

Sensitive features

We consider the following sensitive features:

• Sex
• Marital status
• Age

Among them, sex is binary, marital status is categorical, and age is continuous (requires binning)

Detailed sensitive groups

Sex

• Male
• Female

Marital status

• Married
• Widowed
• Divorced
• Separated
• Never married or under 15 years old

Age

• Less than 30 years
• Between 30 and 40
• Between 40 and 50
• Greater than 50

feature_to_one_hot_encode = "Marital_status" # A sensitive feature. So we will one-hot encode it.
features_unscaled = features_unscaled.replace({feature_to_one_hot_encode : categories[feature_to_one_hot_encode]}) 
features_unscaled = utils.get_one_hot_encoded_df(features_unscaled, [feature_to_one_hot_encode]) 
for key in categories[feature_to_one_hot_encode]:
    val = categories[feature_to_one_hot_encode][key]
    categories[f'{feature_to_one_hot_encode}_' + val] = {1 : 'Yes', 0 : 'No'}
del categories[feature_to_one_hot_encode]

# We consider age as another sensitive feature and consider following age groups
features_unscaled['Age_less_than_30'] = features_unscaled['Age'] < 30
features_unscaled['Age_between_30_and_40'] = (features_unscaled['Age'] >= 30) & (features_unscaled['Age'] < 40)
features_unscaled['Age_between_40_and_50'] = (features_unscaled['Age'] >= 40) & (features_unscaled['Age'] < 50)
features_unscaled['Age_greater_than_50'] = features_unscaled['Age'] >= 50
features_unscaled.drop('Age', axis=1, inplace=True)


# update categories
categories['Age_less_than_30'] = {1 : 'Yes', 0 : 'No'}
categories['Age_between_30_and_40'] = {1 : 'Yes', 0 : 'No'}
categories['Age_between_40_and_50'] = {1 : 'Yes', 0 : 'No'}
categories['Age_greater_than_50'] = {1 : 'Yes', 0 : 'No'}

Scaling dataset

scaler = MinMaxScaler()
features = features_unscaled.copy()
features[features.columns] = scaler.fit_transform(features[features.columns])
features

	Education	Relationship	Disability	Mobility_status	Military_service	Hearing_difficulty	Vision_difficulty	Cognitive_difficulty	Sex	Race	Marital_status_Divorced	Marital_status_Married	Marital_status_Never married or under 15 years old	Marital_status_Separated	Marital_status_Widowed	Age_less_than_30	Age_between_30_and_40	Age_between_40_and_50	Age_greater_than_50
2	0.695652	0.941176	0.0	0.0	1.000000	1.0	1.0	0.0	0.0	0.000	0.0	0.0	1.0	0.0	0.0	0.0	0.0	0.0	1.0
3	0.782609	0.941176	1.0	0.0	0.333333	1.0	1.0	1.0	0.0	0.000	0.0	0.0	1.0	0.0	0.0	1.0	0.0	0.0	0.0
4	0.478261	0.941176	0.0	1.0	1.000000	1.0	1.0	0.0	1.0	0.000	0.0	0.0	1.0	0.0	0.0	1.0	0.0	0.0	0.0
5	0.652174	0.941176	0.0	0.0	1.000000	1.0	1.0	0.0	1.0	0.000	0.0	0.0	1.0	0.0	0.0	0.0	1.0	0.0	0.0
7	0.478261	0.941176	0.0	0.0	1.000000	0.0	1.0	1.0	0.0	0.000	0.0	0.0	1.0	0.0	0.0	0.0	1.0	0.0	0.0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
47758	0.826087	0.000000	1.0	0.0	1.000000	1.0	1.0	1.0	1.0	0.000	0.0	1.0	0.0	0.0	0.0	0.0	1.0	0.0	0.0
47759	0.652174	0.058824	1.0	0.0	1.000000	1.0	1.0	1.0	0.0	0.000	0.0	1.0	0.0	0.0	0.0	0.0	0.0	1.0	0.0
47762	0.782609	0.235294	1.0	0.0	1.000000	1.0	1.0	1.0	0.0	0.000	0.0	0.0	1.0	0.0	0.0	1.0	0.0	0.0	0.0
47765	0.652174	0.117647	1.0	0.0	1.000000	1.0	1.0	1.0	1.0	0.125	1.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	1.0
47771	0.652174	0.000000	1.0	0.0	1.000000	1.0	1.0	1.0	1.0	0.000	0.0	0.0	0.0	1.0	0.0	0.0	1.0	0.0	0.0

23793 rows × 19 columns

Some helper functions

# mapping categories from unscaled to scaled
assert isinstance(scaler, MinMaxScaler)
categories_scaled = {}

for column in features_unscaled.columns:
    categories_scaled[column] = {}
    idx = list(features_unscaled.columns).index(column)
    if(column not in categories):
        continue
    for key in categories[column].keys():
        if(not math.isnan(key)):
            categories_scaled[column][key * scaler.scale_[idx] + scaler.min_[idx]] = categories[column][key]

# helper function to get group statistics from justicia result
def get_group_statistics(justicia_result):

    groups = []
    positive_prediction_probabilities = []
    for i in range(len(justicia_result.sensitive_group_statistics)):
        group_name_per_sensitive_feature = []
        for j in range(len(justicia_result.sensitive_group_statistics[i][0])):
            sensitive_feature = justicia_result.sensitive_group_statistics[i][0][j][0]
            operator = justicia_result.sensitive_group_statistics[i][0][j][1][0]
            value = justicia_result.sensitive_group_statistics[i][0][j][1][1]
            translated_value = None
            if(operator == '=='):
                translated_value = categories_scaled[sensitive_feature][value]
            else:
                translated_value = categories_scaled[sensitive_feature][1 - value]
            group_name_per_sensitive_feature.append(f"{sensitive_feature} = {translated_value}")
        
        groups.append(",\n".join(group_name_per_sensitive_feature))
        positive_prediction_probability = justicia_result.sensitive_group_statistics[i][1]
        positive_prediction_probabilities.append(round(positive_prediction_probability, 3))
        
    return groups, positive_prediction_probabilities

Training

X_train, X_test, y_train, y_test = train_test_split(features, label, test_size=0.4, random_state=0)

model, model_name = LogisticRegression(solver='lbfgs', max_iter=10000), "lr"
# model, model_name = svm.SVC(kernel='linear'), "svm"
model.fit(X_train, y_train.values.ravel())

LogisticRegression(max_iter=10000)

Classification Performance

yhat = model.predict(X_test)
print(classification_report(y_test, yhat, target_names=['0', '1']))

              precision    recall  f1-score   support

           0       0.71      0.43      0.54      3053
           1       0.77      0.92      0.84      6465

    accuracy                           0.76      9518
   macro avg       0.74      0.67      0.69      9518
weighted avg       0.75      0.76      0.74      9518

Formal fairness quantification

We consider three scenarios with different choices of sensitive features

• Sex
• Sex and age
• Sex and marital status

justicia_results = []
for sensitive_feature_set in [['Sex'], ['Sex', 'Age'], ['Sex', 'Marital_status']]:

    metric = Metric(model=model, 
                    data=X_test, 
                    sensitive_attributes=sensitive_feature_set,
                    verbose=False,
                    discretization_bins=10,
                    encoding="best-correlated",
                    dag_threshold=0.6,
                    dag_filename=f"dag_{model_name}_" + "_".join(sensitive_feature_set) +".pkl")
    metric.compute()
    justicia_results.append(metric)



    # results
    print()
    print(f"Sensitive features: {sensitive_feature_set}")
    groups, positive_prediction_probabilities = get_group_statistics(justicia_results[-1])
    plt.rcParams["font.family"] = "serif"
    plt.rcParams['figure.figsize'] = (8, 3)
    fig, ax = plt.subplots()
    bars = ax.barh(groups, positive_prediction_probabilities, 0.5)
    ax.bar_label(bars, fontsize=labelsize)
    plt.title(f"Statistical Parity = {round(justicia_results[-1].statistical_parity_difference, 3)}", fontsize=fontsize)
    plt.xlabel(r"$\Pr[\widehat{Y} = 1 | $" + f"{(', ').join(sensitive_feature_set)}]", fontsize=fontsize)
    # plt.ylabel("Sensitive groups", fontsize=fontsize)
    plt.xticks(fontsize=labelsize)
    plt.yticks(fontsize=labelsize)
    plt.xlim(0.65, 0.9) # careful, this is hardcoded
    plt.tight_layout()
    plt.show()
    plt.clf()

Sensitive features: ['Sex']

Sensitive features: ['Sex', 'Age']

<Figure size 800x300 with 0 Axes>

Sensitive features: ['Sex', 'Marital_status']

<Figure size 800x300 with 0 Axes>

<Figure size 800x300 with 0 Axes>

Explaining the sources of unfairness

fairXplainer = FairXplainer(model,
                            X_test,
                            ['Sex', 'Marital_status'])
fairXplainer.compute(maxorder=2, spline_intervals=3, verbose=False, maxiter=10, cpu_time=2000)

Individual influences

k = 10
fairXplainer_result = fairXplainer.get_weights(k=k)
plt = fif_plot(fairXplainer_result, 
               draw_waterfall=True, 
               labelsize=labelsize,
               fontsize=fontsize,
               figure_size=(9,5), 
               text_y_pad=0.2, 
               xlim=None, 
               result_x_pad=0.02,
               x_label="Influence on Statistical Parity"
        )
plt.tight_layout()
plt.show()
plt.clf()
# print("Exact statistical parity", fairXplainer.statistical_parity_sample())

<Figure size 900x500 with 0 Axes>

Intersectional influences

k = 10
fairXplainer_result = fairXplainer.get_top_k_weights(k=k)

plt = fif_plot(fairXplainer_result, 
               draw_waterfall=True, 
               labelsize=labelsize,
               fontsize=fontsize, 
               figure_size=(9,5), 
               text_y_pad=0.2, 
               xlim=None, 
               result_x_pad=0.01,
               x_label="Influence on Statistical Parity",
        )
plt.tight_layout()
plt.show()
plt.clf()

<Figure size 900x500 with 0 Axes>

Improving fairness by a pre-processing fairness algorithm

encoded_df_train = X_train.copy()
encoded_df_train['target'] = y_train

encoded_df_test = X_test.copy()
encoded_df_test['target'] = y_test


binary_label_dataset_train = aif360.datasets.BinaryLabelDataset(
    favorable_label=0,
    unfavorable_label=1,
    df=encoded_df_train,
    label_names=['target'],
    protected_attribute_names=['Sex'])
binary_label_dataset_test = aif360.datasets.BinaryLabelDataset(
    favorable_label=0,
    unfavorable_label=1,
    df=encoded_df_test,
    label_names=['target'],
    protected_attribute_names=['Sex'])


# Reweighing algorithm
rw = Reweighing(unprivileged_groups=[{'Sex' : 1}], privileged_groups=[{'Sex' : 0}])
rw.fit(binary_label_dataset_train)

binary_label_dataset_trans_train = rw.transform(binary_label_dataset_train)
dataset_transf_train = binary_label_dataset_trans_train.convert_to_dataframe()[0]
dataset_transf_test = rw.transform(binary_label_dataset_test).convert_to_dataframe()[0]

X_trans_train = dataset_transf_train.drop(['target'], axis = 1)
y_trans_train = dataset_transf_train['target']


X_trans_test = dataset_transf_test.drop(['target'], axis = 1)
y_trans_test = dataset_transf_test['target']

Retraining after fairness repair

model_trans, model_trans_name = LogisticRegression(solver='lbfgs', max_iter=10000), "lr"
model_trans.fit(X_trans_train, y_trans_train, sample_weight= binary_label_dataset_trans_train.instance_weights)

LogisticRegression(max_iter=10000)

Compare the difference in statistical parity with respect to Sex

fairXplainer_before = FairXplainer(model, X_test, ['Sex'])
fairXplainer_before.compute(compute_sp_only=True)
statistical_parity_before = round(fairXplainer_before.statistical_parity_sample(), 3)



fairXplainer_after = FairXplainer(model_trans, X_trans_test, ['Sex'])
fairXplainer_after.compute(compute_sp_only=True)
statistical_parity_after = round(fairXplainer_after.statistical_parity_sample(), 3)

print("Statistical Parity Before: ", statistical_parity_before)
print("Statistical Parity After: ", statistical_parity_after)

# bar plot of statistical parity before and after
plt.rcParams["font.family"] = "serif"
plt.rcParams['figure.figsize'] = (8, 3)
fig, ax = plt.subplots()
bars = ax.barh(['Before', 'After'], [statistical_parity_before, statistical_parity_after], 0.5)
ax.bar_label(bars, fontsize=labelsize)
plt.xlabel("Statistical Parity", fontsize=fontsize)
plt.title("Impact of Fairness Repair", fontsize=fontsize)
plt.xticks(fontsize=labelsize)
plt.yticks(fontsize=labelsize)
plt.show()
plt.clf()

Statistical Parity Before:  0.043
Statistical Parity After:  0.009

<Figure size 800x300 with 0 Axes>

Explaining the sources of bias after fairness repair

fairXplainer_trans = FairXplainer(model_trans,
                                  X_trans_test,
                                  ['Sex'])
fairXplainer_trans.compute(maxorder=2, spline_intervals=3, verbose=False, maxiter=10, cpu_time=3000)
# This may take more than 10 minutes

k = 10
fairXplainer_trains_result = fairXplainer_trans.get_top_k_weights(k=k)

plt = fif_plot(fairXplainer_trains_result, 
               draw_waterfall=True, 
               labelsize=labelsize,
               fontsize=fontsize, 
               figure_size=(9,5), 
               text_y_pad=0.2, 
               xlim=None, 
               result_x_pad=0.005,
               x_label="Influence on Statistical Parity"
)
plt.tight_layout()
plt.show()
plt.clf()

<Figure size 900x500 with 0 Axes>