54Walk me through how you run a bias and fairness audit on a deployed model.▼hardMicrosoftGoogleLinkedIn1 replies◆ premiumA fairness audit is a structured process: pick protected attributes, choose metrics that fit the harm, measure disaggregated performance, and document findings. The signal is knowing the metrics conflict and which one the use case requires. Here is the answer.Open full answer →
61Your model passes bias checks for gender and for race separately, but fails for Black women. How do you handle intersectional fairness?▼hardGoogleMicrosoftLinkedIn◆ premiumSingle-axis fairness audits are just another form of averaging, and they average away the exact group that is being harmed. The hard part is not noticing that: it is handling the combinatorial blowup, the tiny cells, and the multiple-testing problem without chasing ghosts.Open full answer →
