It's Time to Optimize Data Algorithms with Fairness Considerations

It may be obvious that the model is biased if it explicitly uses demographic variables. Hiring algorithms may rank females lower because motherhood often disrupts career performance. It is usually difficult to obtain explicit data signals, so modelers turn to proxies that implicitly provide the same signals. A model may just use zip code, but zip code correlates with race and income. P2P lenders commonly develop risk scores based on correlations with neighborhoods, zip codes, stores customers shop at.