Fairness and Bias

The Fairness module provides tools for detecting and mitigating bias in machine learning models. It computes fairness metrics across protected attributes (gender, race, age), identifies disparate impact, and recommends mitigation strategies. The implementation follows established fairness frameworks including demographic parity, equalized odds, and calibration.

Fairness Metrics

Metric	Definition	Threshold
Demographic Parity	Equal positive prediction rate across groups	Ratio above 0.8
Equalized Odds	Equal TPR and FPR across groups	Difference below 0.1
Equal Opportunity	Equal TPR across groups	Difference below 0.1
Calibration	Equal probability accuracy across groups	Difference below 0.05
Disparate Impact	Ratio of positive outcomes between groups	Ratio above 0.8
Predictive Parity	Equal precision across groups	Difference below 0.1

Run Fairness Assessment

POST /api/v1/governance/fairness

{
  "model_id": "model-xyz789",
  "dataset": {
    "source": "sql",
    "query": "SELECT * FROM ml_features.loan_applications"
  },
  "target_column": "approved",
  "protected_attributes": ["gender", "age_group", "ethnicity"],
  "metrics": ["demographic_parity", "equalized_odds", "disparate_impact"],
  "threshold": 0.8
}

Response

{
  "model_id": "model-xyz789",
  "overall_fair": false,
  "groups_analyzed": 3,
  "results": [
    {
      "attribute": "gender",
      "groups": ["male", "female"],
      "metrics": {
        "demographic_parity": {"ratio": 0.92, "pass": true},
        "equalized_odds": {"tpr_diff": 0.05, "fpr_diff": 0.03, "pass": true},
        "disparate_impact": {"ratio": 0.91, "pass": true}
      },
      "fair": true
    },
    {
      "attribute": "age_group",
      "groups": ["18-30", "31-50", "51+"],
      "metrics": {
        "demographic_parity": {"ratio": 0.72, "pass": false},
        "equalized_odds": {"tpr_diff": 0.15, "fpr_diff": 0.08, "pass": false},
        "disparate_impact": {"ratio": 0.74, "pass": false}
      },
      "fair": false,
      "flagged_groups": ["51+"]
    }
  ],
  "recommendations": [
    "Age group '51+' shows significant bias in approval rate",
    "Consider reweighting training data or applying post-processing calibration"
  ]
}

Bias Mitigation Strategies

Strategy	Phase	Method
Reweighting	Pre-processing	Adjust sample weights to equalize group representation
Resampling	Pre-processing	Oversample underrepresented groups
Disparate Impact Remover	Pre-processing	Transform features to remove disparate impact
Adversarial Debiasing	In-processing	Add adversarial fairness constraint during training
Calibrated Equalized Odds	Post-processing	Adjust decision threshold per group
Reject Option Classification	Post-processing	Give favorable outcome to uncertain cases

Apply Mitigation

POST /api/v1/governance/fairness/mitigate

{
  "model_id": "model-xyz789",
  "strategy": "calibrated_equalized_odds",
  "protected_attribute": "age_group",
  "target_metric": "equalized_odds",
  "target_threshold": 0.1
}

Fairness Monitoring

In production, fairness metrics are tracked over time to detect drift in model fairness:

Metric	Frequency	Alert Threshold
Demographic parity ratio	Hourly	Below 0.8
Equalized odds difference	Hourly	Above 0.15
Disparate impact ratio	Daily	Below 0.75
Group-level accuracy	Daily	Divergence above 10%

Configuration

Environment Variable	Default	Description
`FAIRNESS_THRESHOLD`	`0.8`	Default fairness ratio threshold
`FAIRNESS_PROTECTED_ATTRIBUTES`	`gender,age,ethnicity`	Default protected attributes
`FAIRNESS_MONITORING_INTERVAL`	`3600`	Monitoring interval in seconds
`FAIRNESS_ALERT_ENABLED`	`true`	Enable fairness degradation alerts

Explainability (SHAP/LIME)Compliance & Audit