---

name: fairness-bias-auditor description: Evaluates machine learning models for demographic bias using the fairlearn library. Use this skill immediately after training any predictive model.

Fairness & Bias Auditor

Automated decision systems in education can perpetuate inequality. Run this audit after training any predictive model.

Audit Protocol

1. Metric Calculation

Use fairlearn to calculate:

• Demographic Parity Difference: Selection rates across groups
• Equalized Odds: True positive/false positive rates across groups

2. Protected Attributes

For higher education analytics, use:

• PCTPELL (Pell Grant rate) as socioeconomic proxy
• Racial demographics where available

3. Four-Fifths Rule

If selection rate ratio between privileged and unprivileged groups < 0.8, flag as violation.

from fairlearn.metrics import demographic_parity_ratio

ratio = demographic_parity_ratio(
    y_true,
    y_pred,
    sensitive_features=sensitive_group
)
if ratio < 0.8:
    print("WARNING: Four-fifths rule violation detected")

Visualization Requirements

1. Generate disparity plots using CGI color palette
2. Always include the audit results visibly in deliverables
3. Document any mitigations applied (e.g., sample reweighting)

Remediation Options

If bias detected:

1. Adjust sample weights to achieve demographic parity
2. Use fairlearn.reductions.ExponentiatedGradient for constrained optimization
3. Document the trade-off between accuracy and fairness