Sort Bias Signals by Their Likely Source
Sort fairness signals by likely source so the team investigates the right part of the pipeline first.
Sort each fairness signal by the source you should inspect first. Sampling / representation Label / historical process Evaluation / metric choice Deployment / threshold policy A customer tier appears rarely in the training data compared with the population being scored. The target is based on past manager promotion decisions that may reflect unequal sponsorship access. Overall accuracy is strong, but no one has checked false negatives by subgroup. One global threshold denies benefits to a group more often even though score distributions differ. Applicants from alternative education programs barely appear in the historical sample. A complaint model is trained…
Sign up free — one personalized lesson every day, matched to your role and goals.
Already have an account? Sign in