Individual Fairness

Explanation

Individual fairness checks the principle that similar people should be treated similarly — two loan or insurance applicants with near-identical relevant profiles should receive near-identical decisions regardless of incidental attributes. It matters because group-level parity can still hide cases where look-alike customers get inconsistent outcomes, undermining trust and fair-lending obligations. It is measured by Similarity-based Decision Consistency, computed as Decision Consistency = 1 − [Σ(outcome(x_i) − outcome(x_j)) / d(x_i, x_j)] / n, where d(x_i, x_j) is the distance between similar customers; the more two near-identical customers diverge in outcome relative to their small distance, the lower the consistency. To implement it, define a domain-justified similarity/distance metric over meaningful features, pair up close customers from validation and production samples, compute outcome differences scaled by distance, and log the per-pair contributions and aggregate score as evidence on a fairness dashboard. Run it pre-deployment and periodically in monitoring, especially after model updates. The threshold requires at least 90% consistency for similar profiles; if consistency drops below 90% it triggers a breach — investigate the inconsistent pairs, check for unstable features or proxies, remediate, and record sign-off, supporting RBI Fairness and Equity and NIST MAP 2.1/4.1 expectations.