Examples · Deciding

decide

The score-only path versus the explained payload, and how missing values are handled under each policy.

Code#

decide.py

from _setup import artifact, row
 
from compileml.runtime import decide
 
fast = decide(artifact, row, explain=False)
print("score only :", {k: fast[k] for k in ("band", "pd", "latent_int")})
 
full = decide(artifact, row, explain=True, top_k=3)
print()
print("adverse reasons, largest first:")
for reason in full["reasons_negative"]:
    print(f"  {reason['impact_int']:>5}  {reason['code']:<22} {reason['message']}")
 
# A missing value is imputed from features.baseline under this policy.
patched = list(row)
patched[0] = None
print()
print("with LIMIT_BAL missing:", decide(artifact, patched, explain=False)["band"])
print("with LIMIT_BAL present:", fast["band"])

Output#

Captured from an actual run against compileml 0.9.0 and the UCI credit panel. If this script stops working, the build fails.

captured in CIdecide.py

score only : {'band': 'G09', 'pd': 0.378049, 'latent_int': 341}
 
adverse reasons, largest first:
    117  REPAYMENT_STATUS_M1    Repayment was behind schedule 1 month(s) before assessment.
     84  REPAYMENT_STATUS_M2    Repayment was behind schedule 2 month(s) before assessment.
     10  AMOUNT_PAID_M4         The amount paid 4 month(s) before assessment was small.
 
with LIMIT_BAL missing: G09
with LIMIT_BAL present: G09

Notes#

explain=False is the sub-millisecond path. explain=True adds O(n_features²) traversals.
reject

API reference: decide →

Related notebook Spec §8 Missing values