LOADING FAIRSWARM
LOADING FAIRSWARM
Fairness Documentation
Understand fairness metrics in plain language, see formal definitions, and connect each term to practical deployment risks in India.
Plain English: Compares favorable outcomes between protected and reference groups.
DI = P(y=1 | unprivileged) / P(y=1 | privileged)
Example: If loan approvals are 40% for Group A and 20% for Group B, DI is 0.50.
Why it matters in India: In lending, applicants from historically underserved castes or districts can be disproportionately denied even with similar repayment history.
Plain English: Difference in positive prediction rates across groups.
SPD = P(y=1 | unprivileged) - P(y=1 | privileged)
Example: If hiring model selects 28% women and 35% men, SPD is -0.07.
Why it matters in India: Hiring models can under-select women or candidates from non-metro regions despite similar qualifications.
Plain English: Checks if true positive rates are similar across groups.
EOD = TPR(unprivileged) - TPR(privileged)
Example: Approved truly eligible loan applicants should have similar TPR across groups.
Why it matters in India: Public welfare scoring systems should not miss eligible candidates from rural or tribal communities at higher rates.
Plain English: Ensures both true positive and false positive rates are balanced.
Compare TPR and FPR across protected groups
Example: A fraud model that flags one region more often despite similar behavior violates equalized odds.
Why it matters in India: Risk and compliance models for digital payments should not penalize specific states due to historical data skew.
Plain English: Checks whether precision is similar across groups.
PPV(unprivileged) ~= PPV(privileged)
Example: When the model predicts approval, quality should be equally reliable for all groups.
Why it matters in India: In hiring and scholarships, predictions must maintain comparable reliability across gender, caste, and region.
Enter your ratio to quickly check whether it passes common fairness thresholds.
Typical compliance checks use 0.80 as the minimum acceptable ratio.
Explore practical frameworks and references to reduce bias during model lifecycle reviews.