boundary_fragility
def boundary_fragility(decisions, protected, threshold_int: int, *, labels=None, near: int = 25) -> dict
from compileml.fairness import boundary_fragility
§9 — how close each group sits to the cutoff.
An adverse impact ratio is a snapshot. This is its derivative: a group clustered near the boundary is one policy tweak away from a disparate outcome even when today's ratio passes comfortably, and that exposure is invisible to every outcome-level test.
Parameters#
| Name | Type | Default | Kind |
|---|---|---|---|
| decisions | — | required | positional |
| protected | — | required | positional |
| threshold_int | int | required | positional |
| labels | — | None | keyword-only |
| near | int | 25 | keyword-only |
Returns#
dict