Austin Li — Projects

While working at Zest AI, I developed an open-source machine learning algorithm for race and ethnicity estimation using full name and home address as inputs. The goal was to improve upon the most widely used racial and ethnic estimation method — Bayesian Improved Surname Geocoding (BISG), developed by RAND Corporation in 2009 — to better analyze racial equity and outcomes in healthcare, finance, and other cases where race or ethnicity data need to be imputed.

Built with gradient boosting, trained on voter registration data from North Carolina and Florida, validated on a national sample using adjusted tract-level American Community Survey (ACS) data. Found that compared to BISG, the zrp correctly identifies 25% more African Americans as African American, 35% fewer African Americans as non-African American, and 60% fewer White Americans as non-White.

Links: GitHub