FairnessAudit

from _setup import SEED, X_test, artifact, feature_names, y_test
from sklearn.model_selection import train_test_split
 
from compileml.datasets import load_credit_default
from compileml.fairness import FairnessAudit
from compileml.runtime import decide
 
# The artifact was compiled without demographics. Load them separately and
# repeat the same split, so each holdout row gets its own SEX value back.
X_demo, y_demo, demo_names = load_credit_default(include_demographics=True)
_, X_demo_test, _, _ = train_test_split(
    X_demo, y_demo, test_size=0.25, stratify=y_demo, random_state=SEED
)
sex = X_demo_test[:, demo_names.index("SEX")].astype(int).tolist()
assert "SEX" not in feature_names
 
# §6 and §8 need the per-feature contributions, which decide() leaves out
# by default.
decisions = [decide(artifact, r.tolist(), include_contributions=True) for r in X_test]
 
# The cutoff is a policy, not a statistic: here, approve bands G01–G07.
bands = artifact["bands"]
cutoff = bands["edges_int"][bands["labels"].index("G08")]
 
audit = FairnessAudit(
    decisions,
    y_test,
    sex,
    labels={1: "Male", 2: "Female"},
    threshold_int=cutoff,
    artifact=artifact,
    X=X_test,
    protected_feature="SEX",  # named, so §11 can say whether the model uses it
)
audit.compute_all()
audit.print_summary()
 
# The decomposition is exact because each decision's contributions sum to its
# score's distance from the baseline, in integers. Summed over a group, the
# identity still holds, so the per-feature gaps sum to the group gap.
for code, name in ((1, "Male"), (2, "Female")):
    rows = [d for d, s in zip(decisions, sex) if s == code]
    movement = sum(2 * (d["raw_micro"] - d["baseline_micro"]) for d in rows)
    parts = sum(c["impact_half_micro"] for d in rows for c in d["contributions"])
    print(f"{name:<7} movement {movement:>13,}  contributions {parts:>13,}  equal {movement == parts}")
 
print()
print("§11 counterfactual:", audit.section(11)["status"])