Fairness & Inclusiveness

# pip install fairlearn aif360 scikit-learn pandas numpy matplotlib

# ─── 1. Load and Inspect Data for Bias ───────────────────────
import pandas as pd
import numpy as np
from sklearn.model_selection import train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.preprocessing import StandardScaler

# Adult Income dataset (classic fairness benchmark)
from sklearn.datasets import fetch_openml
data = fetch_openml("adult", version=2, as_frame=True)
df = data.frame.copy()

# Target: income >50K (1) or <=50K (0)
df['income_binary'] = (df['class'] == '>50K').astype(int)

# Sensitive attribute: sex
df['sex_binary'] = (df['sex'] == 'Male').astype(int)  # 1=Male, 0=Female

features = ['age', 'educational-num', 'hours-per-week', 'capital-gain', 'capital-loss']
X = df[features].fillna(0)
y = df['income_binary']
sensitive = df['sex_binary']

X_train, X_test, y_train, y_test, s_train, s_test = train_test_split(
    X, y, sensitive, test_size=0.3, random_state=42, stratify=y)

scaler = StandardScaler()
X_train_scaled = scaler.fit_transform(X_train)
X_test_scaled = scaler.transform(X_test)

# ─── 2. Train Baseline Model ──────────────────────────────────
model = LogisticRegression(max_iter=1000, random_state=42)
model.fit(X_train_scaled, y_train)
y_pred = model.predict(X_test_scaled)
y_prob = model.predict_proba(X_test_scaled)[:, 1]

# ─── 3. Compute Fairness Metrics ─────────────────────────────
from fairlearn.metrics import (
    MetricFrame,
    demographic_parity_difference,
    demographic_parity_ratio,
    equalized_odds_difference,
    selection_rate,
    true_positive_rate,
    false_positive_rate,
)
from sklearn.metrics import accuracy_score, precision_score, recall_score

metrics = {
    "accuracy": accuracy_score,
    "precision": precision_score,
    "recall": recall_score,
    "selection_rate": selection_rate,
    "true_positive_rate": true_positive_rate,
    "false_positive_rate": false_positive_rate,
}

mf = MetricFrame(
    metrics=metrics,
    y_true=y_test,
    y_pred=y_pred,
    sensitive_features=s_test,
)

print("Fairness Metrics by Gender:")
print(mf.by_group.to_string())
print(f"\nDemographic Parity Difference: {demographic_parity_difference(y_test, y_pred, sensitive_features=s_test):.4f}")
print(f"Demographic Parity Ratio:       {demographic_parity_ratio(y_test, y_pred, sensitive_features=s_test):.4f}")
print(f"Equalized Odds Difference:      {equalized_odds_difference(y_test, y_pred, sensitive_features=s_test):.4f}")

# Check 80% rule (EEOC Disparate Impact)
sel_male   = y_pred[s_test == 1].mean()
sel_female = y_pred[s_test == 0].mean()
dir_ratio = sel_female / sel_male
print(f"\nSelection rates: Male={sel_male:.3f}, Female={sel_female:.3f}")
print(f"Disparate Impact Ratio: {dir_ratio:.3f} {'(PASS ≥0.8)' if dir_ratio >= 0.8 else '(FAIL <0.8 — potential discrimination)'}")

# ─── 4. Pre-processing Mitigation: Reweighting ────────────────
from aif360.datasets import BinaryLabelDataset
from aif360.algorithms.preprocessing import Reweighing

# Prepare AIF360 dataset
df_aif = pd.DataFrame(X_train, columns=features)
df_aif['income'] = y_train.values
df_aif['sex'] = s_train.values

aif_dataset = BinaryLabelDataset(
    df=df_aif,
    label_names=['income'],
    protected_attribute_names=['sex'],
)

# Reweighing: assign instance weights to equalize representation
RW = Reweighing(unprivileged_groups=[{'sex': 0}],
                privileged_groups=[{'sex': 1}])
RW.fit(aif_dataset)
rw_dataset = RW.transform(aif_dataset)
sample_weights = rw_dataset.instance_weights

model_rw = LogisticRegression(max_iter=1000, random_state=42)
model_rw.fit(X_train_scaled, y_train, sample_weight=sample_weights)
y_pred_rw = model_rw.predict(X_test_scaled)

print("\nAfter Reweighing:")
dpd_rw = demographic_parity_difference(y_test, y_pred_rw, sensitive_features=s_test)
print(f"  Demographic Parity Difference: {dpd_rw:.4f} (was {demographic_parity_difference(y_test, y_pred, sensitive_features=s_test):.4f})")
print(f"  Accuracy: {accuracy_score(y_test, y_pred_rw):.4f} (was {accuracy_score(y_test, y_pred):.4f})")

# ─── 5. Post-processing Mitigation: Threshold Optimization ────
from fairlearn.postprocessing import ThresholdOptimizer
from fairlearn.reductions import ExponentiatedGradient, DemographicParity

# Exponentiated Gradient (in-processing): fair constraint during training
constraint = DemographicParity()
mitigator = ExponentiatedGradient(
    LogisticRegression(max_iter=1000),
    constraints=constraint,
)
mitigator.fit(X_train_scaled, y_train, sensitive_features=s_train)
y_pred_eg = mitigator.predict(X_test_scaled)

print("\nAfter Exponentiated Gradient (DemographicParity constraint):")
dpd_eg = demographic_parity_difference(y_test, y_pred_eg, sensitive_features=s_test)
print(f"  Demographic Parity Difference: {dpd_eg:.4f}")
print(f"  Accuracy: {accuracy_score(y_test, y_pred_eg):.4f}")

# ─── 6. Fairness Report (ISO 42001 A.6.2.3 evidence) ─────────
print("\n" + "="*60)
print("FAIRNESS AUDIT REPORT — ISO 42001 Control A.6.2.3")
print("="*60)
print(f"Model: Logistic Regression | Dataset: Adult Income (UCI)")
print(f"Sensitive Attribute: sex (Male=1, Female=0)")
print(f"\nBefore mitigation:")
print(f"  Accuracy: {accuracy_score(y_test, y_pred):.4f}")
print(f"  DPD: {demographic_parity_difference(y_test, y_pred, sensitive_features=s_test):.4f}")
print(f"  DIR: {dir_ratio:.4f} {'PASS' if dir_ratio >= 0.8 else 'FAIL (requires mitigation)'}")
print(f"\nAfter mitigation (Exponentiated Gradient):")
print(f"  Accuracy: {accuracy_score(y_test, y_pred_eg):.4f}")
print(f"  DPD: {dpd_eg:.4f}")
print(f"\nRecommendation: {'Continue to monitoring' if abs(dpd_eg) < 0.05 else 'Further review required'}")