Where: explainers/fairness-through-unawareness.md. The prose claims the Tenant Screening core features (Supervision_Risk_Score_First, prior-arrest counts, Percent_Days_Employed) "carry real correlation with race" strongly enough to explain why dropping race increased the demographic-parity gap.
The gap: running the file's own unawareness_gap_check() on exactly those features against the real Tenant Screening/tenant-screening-data.csv:
unawareness_gap_check(df, features=['Supervision_Risk_Score_First', ...], protected_col='Race')
gives a reconstruction accuracy of only 0.602 (0.617 with the fuller feature list the prose names) - barely above the 0.575 majority-class base rate, and well below the function's own documented 0.75 "reconstructs well" threshold. Run against its own supplied features and its own supplied threshold, the tool this article provides would verdict this as "poorly predicts the protected attribute" - the opposite of what the surrounding narrative claims it shows.
Fix direction: either correct the narrative to match what the tool actually reports for this feature set (the features don't reconstruct race well, so that's not why the gap moved - some other mechanism is), or find and cite a feature combination that genuinely clears the 0.75 threshold if that's the point being made.
Where:
explainers/fairness-through-unawareness.md. The prose claims the Tenant Screening core features (Supervision_Risk_Score_First, prior-arrest counts,Percent_Days_Employed) "carry real correlation with race" strongly enough to explain why dropping race increased the demographic-parity gap.The gap: running the file's own
unawareness_gap_check()on exactly those features against the realTenant Screening/tenant-screening-data.csv:gives a reconstruction accuracy of only 0.602 (0.617 with the fuller feature list the prose names) - barely above the 0.575 majority-class base rate, and well below the function's own documented 0.75 "reconstructs well" threshold. Run against its own supplied features and its own supplied threshold, the tool this article provides would verdict this as "poorly predicts the protected attribute" - the opposite of what the surrounding narrative claims it shows.
Fix direction: either correct the narrative to match what the tool actually reports for this feature set (the features don't reconstruct race well, so that's not why the gap moved - some other mechanism is), or find and cite a feature combination that genuinely clears the 0.75 threshold if that's the point being made.