Getting Started with vfairness

# Core (numpy, pandas, scipy)
pip install vfairness

# With visualization support (matplotlib, plotly)
pip install vfairness[viz]

# With interactive dashboards (Dash)
pip install vfairness[reporting]

# Full installation (all optional dependencies)
pip install vfairness[all]

import vfairness
print(f"vfairness {vfairness.__version__}")

import numpy as np
from vfairness import FairnessAnalyzer, demographic_parity_difference

# Sample data
y_true = np.array([1, 0, 1, 0, 1, 0, 1, 0, 1, 1] * 100)
y_pred = np.array([1, 0, 1, 1, 1, 0, 0, 0, 1, 1] * 100)
gender = np.array(['M','M','M','M','M','F','F','F','F','F'] * 100)

# Single metric
dp = demographic_parity_difference(y_true, y_pred, gender)
print(f"Demographic Parity Difference: {dp:.3f}")

# Full analyzer with confidence intervals
analyzer = FairnessAnalyzer(y_true, y_pred, gender, min_group_size=30)
result = analyzer.demographic_parity_difference(
    include_ci=True, n_bootstrap=5000, confidence_level=0.95
)
print(f"DP Diff: {result.value:.3f}  95% CI: [{result.ci_lower:.3f}, {result.ci_upper:.3f}]")

# All metrics at once
for name, value in analyzer.compute_all_metrics().items():
    print(f"  {name}: {value:.3f}")

from vfairness.llm import OutputAnalyzer

analyzer = OutputAnalyzer(alpha=0.05)
result = analyzer.analyze_sentiment(texts_a, texts_b)

# Every result has audit metadata
print(result.metadata.timestamp)        # ISO 8601 UTC
print(result.metadata.library_version)  # e.g. "0.1.0"
print(result.metadata.parameters)       # {"alpha": 0.05, "metric": "sentiment", ...}

# Serialize for database storage
data = result.to_dict()    # Plain dict
json_str = result.to_json()  # JSON string

# Structured logging (configure once)
import logging
logging.basicConfig(level=logging.INFO)
# All vfairness operations now emit structured logs:
# INFO:vfairness.llm.output_analysis:Starting analyze_all with 4 metrics, 50 samples per group
# WARNING:vfairness.llm.output_analysis:Sample size (18) below recommended minimum of 25

from tqdm import tqdm

# With tqdm progress bar
pbar = tqdm(total=100)
proxy.send_batch(
    prompts=my_prompts,
    n_runs=25,
    progress_callback=lambda current, total: pbar.update(1)
)
pbar.close()

import pandas as pd
from vfairness import BiasDetector, detect_historical_patterns, identify_proxy_variables

# Full audit
detector = BiasDetector(df, target_column='approved', protected_attributes=['gender', 'age'])
report = detector.full_audit(
    protected_columns=['gender'],
    population_benchmarks={'gender': {'M': 0.49, 'F': 0.51}}
)
print(report.summary())

# Historical discrimination patterns
for r in detect_historical_patterns(df, column_names=df.columns.tolist()):
    print(f"  {r.feature}: {r.risk_level.value} - {r.pattern_type}")

# Proxy variable detection
for p in identify_proxy_variables(df, protected_columns=['gender'], correlation_threshold=0.3):
    print(f"  {p.feature} -> {p.protected_attribute}  (r={p.correlation:.3f})")

from vfairness.post_processing.calibration import CalibrationAnalyzer

analyzer = CalibrationAnalyzer(
    y_true=labels, y_prob=model_probabilities,
    protected_attr=gender, attribute_name='gender'
)

report = analyzer.full_analysis(context='lending')
print(f"Well Calibrated: {report.is_well_calibrated}")
print(f"Significant Disparity: {report.has_significant_disparity}")
print(f"Overall ECE: {report.overall_metrics['ece']:.4f}")

# Apply group-specific recalibration
analyzer.fit_calibrator(method='isotonic')
calibrated_probs = analyzer.transform(y_prob_test, gender_test)

from vfairness import FairnessAnalyzer

# FairExplAIner mode for per-metric explanations
analyzer = FairnessAnalyzer(y_true, y_pred, gender, fair_explainer=True)
report = analyzer.get_report(include_ci=True)

for name, expl in report['explanations']['metrics'].items():
    print(f"{name}: {expl['severity']} - {expl['evaluation']}")

# FairnessExplainer for cross-module explanations
from vfairness.explainer import FairnessExplainer

explanation = FairnessExplainer.explain(audit_report)  # Works for any result type
print(explanation.severity)           # 'info' | 'low' | 'medium' | 'high' | 'critical'
print(explanation.recommendations)    # Actionable next steps

from vfairness import classification_fairness_report, plot_fairness_metrics, create_fairness_dashboard
from vfairness.rendering import radar_chart_to_svg, fairness_detailed_report_to_svg

report = classification_fairness_report(y_true, y_pred, gender, include_ci=True)

# Matplotlib (static)
fig = plot_fairness_metrics(report, style='academic', show_thresholds=True)
fig.savefig('fairness_metrics.png', dpi=300, bbox_inches='tight')

# Plotly (interactive)
dashboard = create_fairness_dashboard(report, style='modern')
dashboard.write_html('fairness_dashboard.html')

# SVG templates (zero-dependency)
svg = radar_chart_to_svg(metrics_data, save_path='radar.svg')

from vfairness.operations.monitoring import FairnessMonitor, FairnessDriftDetector

# Real-time monitoring
monitor = FairnessMonitor(window_size=1000)
window = monitor.update_and_check(predictions, labels, sensitive_attrs)
if window.any_alert:
    print(f"Alert! Flagged metrics: {[k for k, v in window.alerts.items() if v]}")

# Multi-scale drift detection
detector = FairnessDriftDetector()
drift = detector.check_drift(reference_data, current_data, metric="demographic_parity")
print(f"Drift detected: {drift.drift_detected}  (score: {drift.overall_drift_score:.3f})")

from vfairness.operations.reporting import MetricsStore, FairnessDashboard, ReportGenerator

store = MetricsStore()
store.ingest_from_monitor(monitor)

# Interactive Plotly dashboard
dashboard = FairnessDashboard(store)
fig = dashboard.create_executive_view()
fig.write_html("dashboard.html")

# Automated multi-format report (HTML, PDF, JSON)
gen = ReportGenerator(store, dashboard)
report = gen.generate_executive_report()
report.save("executive_report.html")

from vfairness.operations.experimentation import (
    FairnessExperiment, FairnessPowerAnalyzer, ExperimentAnalysis
)

exp = FairnessExperiment(
    control_data=df_ctrl, treatment_data=df_treat,
    protected_attributes=['gender', 'race'], outcome_column='approved',
)
result = exp.run_full_analysis()

# Per-intersection power analysis
power = FairnessPowerAnalyzer(exp)
print(power.get_power_summary())

# Multi-objective analysis and deployment decision
analysis = ExperimentAnalysis(result, experiment=exp)
rec = analysis.decision_recommendation()
print(f"Decision: {rec.decision}")  # DEPLOY | HOLD | REVERT

from vfairness.operations.cicd import (
    ModelFairnessGate, HierarchicalGateConfig, FairnessReportCard
)

# Hierarchical gate: overall → single-attribute → intersectional
gate = ModelFairnessGate(
    metrics=['demographic_parity_difference', 'equal_opportunity_difference'],
    thresholds={'demographic_parity_difference': 0.1, 'equal_opportunity_difference': 0.1}
)
config = HierarchicalGateConfig(
    check_intersections=True, intersection_depth=2
)
decision = gate.evaluate_hierarchical(
    y_true, y_pred, protected_attributes=['gender', 'race'],
    hierarchical_config=config
)

# Generate PR-ready report card
card = FairnessReportCard(decision, model_name='loan-approval-v2.1')
print(card.to_markdown())  # Paste into PR comment

from vfairness import log_fairness_to_wandb, auto_log_fairness

# W&B experiment tracking
log_fairness_to_wandb(analyzer, prefix='fairness')

# Auto-log decorator — wraps any training function
@auto_log_fairness(backend='mlflow')
def train_model(X, y, sensitive):
    model = LogisticRegression().fit(X, y)
    return model.predict(X)  # fairness metrics logged automatically