reason_parity
def reason_parity(decisions, protected, *, labels=None, top_n: int = 5) -> dict
from compileml.fairness import reason_parity
§10 — are the same reasons cited to both groups?
Regulatory rather than statistical: this is about what applicants are *told*. A systematic difference in cited reasons is a finding even when approval rates match exactly.
Note what is being counted. reasons_negative is the adverse-action
code the applicant would actually be sent, produced by the same runtime
that produced the decision — not a feature-importance ranking standing in
for one. The audit is of disclosed reasons, not of a proxy for them.
Parameters#
| Name | Type | Default | Kind |
|---|---|---|---|
| decisions | — | required | positional |
| protected | — | required | positional |
| labels | — | None | keyword-only |
| top_n | int | 5 | keyword-only |
Returns#
dict