Prof. Ayelet Israeli from Harvard University (US) was our second guest in the Research Seminar Series this semester. She presented her work on “BEAT Unintended Bias in Personalized Policies”.

Algorithmic bias can affect marketing strategies by disproportionally targeting potential customers from certain groups (e.g., based on their demographic characteristics). Prof. Israeli and her co-workers found that this unintended discrimination is often the result of some attributes of the data correlating with “protected” characteristics, such as gender or race. They propose BEAT (Bias-Eliminating Adapted Trees), which audits the personalization algorithm. BEAT ensures a balanced allocation of individuals across all targeting groups by extracting from the data only those differences, which are not related to any protected attributes. In her paper, Prof. Israeli was able to validate BEAT using both an experiment and a simulation, showing that only removing protected attributes from the data does not solve the discrimination problem, while using BEAT makes the high-dimensional data clearer and allows for non-discriminative resource allocation.

Thank you, Ayelet, for coming to visit us!