On the application and impact of ε-DP and fairness in ambulance engagement time prediction

Selene Cerna, Catuscia Palamidessi

Keywords: Privacy preserving Machine Learning, Differential Privacy, Fairness, Firemen

TL;DR: On the application and impact of differential privacy and fairness mechanisms when predicting ambulance engagement time with RandomForestClassifier

Abstract: This study elaborates on a complete pipeline for the development of a private and fair Machine Learning (ML) model to predict ambulance engagement time. It was shown that sensitive variables reduced their impact on model building with Random Forest as the differential privacy budget (ε) decreased with the GRR and Geometric mechanisms. Also, the application of the Reweighing fairness mechanism negatively affected fairness in private models. Finally, it is possible to keep firefighters' and victims' privacy, recovering an ML model with good performance.