Fair Active Learning For Personalized Policies

As AI increasingly takes on tasks that previously required human intelligence, concerns about biases in algorithms have grown. This thesis proposes a new active learning algorithm to design fair experiments, which can decrease the risk that targeting policies optimized based on these experiments discriminate against protected groups. Unlike other approaches, Fair Active Learning (FAL) tackles fairness during the data acquisition stage and thus mitigates the risk that a model would learn from historical disparities. In addition, FAL focuses on ensuring equal opportunity rather than statistical parity, which allows companies to not compromise on the return of investments of their targeting policies. The algorithm is evaluated on a set of simulations as well as on a real cross-selling campaign. Overall, the results show that FAL improves equal opportunities without sacrificing model accuracy.