Data Loading and Packages Installation

First, install the holisticai package if you haven’t already:

!pip install holisticai[all]

Then, import the necessary libraries.

import warnings

import pandas as pd
from holisticai.bias.metrics import classification_bias_metrics
from holisticai.datasets import load_dataset
from holisticai.pipeline import Pipeline
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler

warnings.filterwarnings("ignore")

Dataset loading

dataset = load_dataset('adult', protected_attribute='sex')
train_test = dataset.train_test_split(test_size=0.2, random_state=42)

train = train_test['train']
test = train_test['test']

dataset

[-] [Dataset]

Instances: 45222

Features: X , y , p_attrs , group_a , group_b

Metadata: sex: {'group_a': 'Male', 'group_b': 'Female'}

1 . Calibrated Equalized Odds

Traditional Implementation

# Define postprocessing model
from holisticai.bias.mitigation import CalibratedEqualizedOdds

mitigator = CalibratedEqualizedOdds(cost_constraint="fnr")
mitigator

[-] [CalibratedEqualizedOdds]

CalibratedEqualizedOdds(cost_constraint=fnr, alpha=None)

---

Type: Bias Mitigation Postprocessing

model = LogisticRegression()

# Standardize data and fit model
scaler = StandardScaler()
X_train = scaler.fit_transform(train['X'])
model.fit(X_train, train['y'])

# Fit postprocessing model
y_pred = model.predict_proba(X_train)
mitigator.fit(train['y'], y_pred, group_a=train['group_a'], group_b=train['group_b'])

# Predict and mitigate bias
X_test = scaler.transform(test['X'])
y_pred = model.predict_proba(X_test)
y_pred = mitigator.transform(test['y'], y_pred, test['group_a'], test['group_b'])["y_pred"]

# Evaluate bias metrics
metrics = classification_bias_metrics(test['group_a'], test['group_b'], y_pred, test['y'], metric_type='both')
metrics

	Value	Reference
Metric
Statistical Parity	0.106757	0
Disparate Impact	2.374672	1
Four Fifths Rule	0.421111	1
Cohen D	0.302012	0
2SD Rule	13.301377	0
Equality of Opportunity Difference	-0.078878	0
False Positive Rate Difference	0.050882	0
Average Odds Difference	-0.013998	0
Accuracy Difference	-0.147467	0

2. Equalized Odds

Traditional Implementation

# Define postprocessing model
from holisticai.bias.mitigation import EqualizedOdds

mitigator = EqualizedOdds(solver='highs', seed=42)
mitigator

[-] [EqualizedOdds]

EqualizedOdds(solver=highs, seed=42)

---

Type: Bias Mitigation Postprocessing

# Standardize data and fit model
scaler = StandardScaler()
X_train = scaler.fit_transform(train['X'])
model.fit(X_train, train['y'])
y_pred = model.predict(X_train)

# Fit postprocessing model
mitigator.fit(train['y'], y_pred, group_a=train['group_a'], group_b=train['group_b'])

# Predict and mitigate bias
X_test = scaler.transform(test['X'])
y_pred = model.predict(X_test)
y_pred = mitigator.transform(y_pred, test['group_a'], test['group_b'])["y_pred"]

# Evaluate bias metrics
metrics = classification_bias_metrics(test['group_a'], test['group_b'], y_pred, test['y'], metric_type='both')
metrics

	Value	Reference
Metric
Statistical Parity	0.089406	0
Disparate Impact	1.622222	1
Four Fifths Rule	0.616439	1
Cohen D	0.222988	0
2SD Rule	9.864874	0
Equality of Opportunity Difference	0.004214	0
False Positive Rate Difference	-0.003179	0
Average Odds Difference	0.000518	0
Accuracy Difference	-0.063250	0

3. LP Debiaser

Traditional Implementation

# Define postprocessing model
from holisticai.bias.mitigation import LPDebiaserBinary

mitigator = LPDebiaserBinary()
mitigator

[-] [LPDebiaserBinary]

LPDebiaserBinary(constraint=EqualizedOdds)

---

Type: Bias Mitigation Postprocessing

model = LogisticRegression()

# Standardize data and fit model
scaler = StandardScaler()
X_train = scaler.fit_transform(train['X'])
model.fit(X_train, train['y'])
y_proba = model.predict_proba(X_train)

# Fit postprocessing model
mitigator.fit(y=train['y'], y_proba=y_proba, group_a=train['group_a'], group_b=train['group_b'])

# Predict and mitigate bias
X_test = scaler.transform(test['X'])
y_pred = model.predict_proba(X_test)
y_pred = mitigator.transform(y_proba=y_pred, group_a=test['group_a'], group_b=test['group_b'])["y_pred"]

# Evaluate bias metrics
metrics = classification_bias_metrics(test['group_a'], test['group_b'], y_pred, test['y'], metric_type='both')
metrics

	Value	Reference
Metric
Statistical Parity	0.110736	0
Disparate Impact	1.359246	1
Four Fifths Rule	0.735702	1
Cohen D	0.229077	0
2SD Rule	10.131199	0
Equality of Opportunity Difference	-0.006369	0
False Positive Rate Difference	0.000221	0
Average Odds Difference	-0.003074	0
Accuracy Difference	0.011336	0

4. ML Debiaser

Traditional Implementation

# Define postprocessing model
from holisticai.bias.mitigation import MLDebiaser

mitigator = MLDebiaser()
mitigator

[-] [MLDebiaser]

MLDebiaser(gamma=1.0, eps=0, eta=0.5, sgd_steps=10000, ...)

---

Type: Bias Mitigation Postprocessing

model = LogisticRegression()

# Standardize data and fit model
scaler = StandardScaler()
X_train = scaler.fit_transform(train['X'])
model.fit(X_train, train['y'])
y_pred = model.predict_proba(X_train)

# Fit postprocessing model
mitigator.fit(y_pred, group_a=train['group_a'], group_b=train['group_b'])

# Predict and mitigate bias
X_test = scaler.transform(test['X'])
y_pred = model.predict_proba(X_test)
y_pred = mitigator.transform(y_pred, test['group_a'], test['group_b'])["y_pred"]

# Evaluate bias metrics
metrics = classification_bias_metrics(test['group_a'], test['group_b'], y_pred, test['y'], metric_type='both')
metrics

[elapsed time: 00:00:00 | iter:2/2]

	Value	Reference
Metric
Statistical Parity	0.078856	0
Disparate Impact	1.503264	1
Four Fifths Rule	0.665219	1
Cohen D	0.194378	0
2SD Rule	8.610334	0
Equality of Opportunity Difference	-0.160993	0
False Positive Rate Difference	0.002258	0
Average Odds Difference	-0.079367	0
Accuracy Difference	-0.084606	0

5. Reject Option

Traditional Implementation

# Define postprocessing model
from holisticai.bias.mitigation import RejectOptionClassification

mitigator = RejectOptionClassification(metric_name="Statistical parity difference")
mitigator

[-] [RejectOptionClassification]

RejectOptionClassification(low_class_thresh=0.01, high_class_thresh=0.99, num_class_thresh=100, num_ROC_margin=50, ...)

---

Type: Bias Mitigation Postprocessing

model = LogisticRegression()

# Standardize data and fit model
scaler = StandardScaler()
X_train = scaler.fit_transform(train['X'])
model.fit(X_train, train['y'])
y_pred = model.predict_proba(X_train)

# Fit postprocessing model
mitigator.fit(train['y'], y_pred, group_a=train['group_a'], group_b=train['group_b'])

# Predict and mitigate bias
X_test = scaler.transform(test['X'])
y_pred = model.predict(X_test)
y_proba = model.predict_proba(X_test)
y_pred = mitigator.transform(y_pred, y_proba, test['group_a'], test['group_b'])["y_pred"]

# Evaluate bias metrics
metrics = classification_bias_metrics(test['group_a'], test['group_b'], y_pred, test['y'], metric_type='both')
metrics

	Value	Reference
Metric
Statistical Parity	0.038357	0
Disparate Impact	2.475245	1
Four Fifths Rule	0.404000	1
Cohen D	0.173384	0
2SD Rule	7.686846	0
Equality of Opportunity Difference	-0.026171	0
False Positive Rate Difference	0.006603	0
Average Odds Difference	-0.009784	0
Accuracy Difference	-0.161902	0