Evaluating Fairness in Predictive Policing Using Domain Knowledge

Abstract

As an increasing number of Artificial Intelligence (AI) systems are ingrained in our day-to-day lives, it is crucial that they are fair and trustworthy. Unfortunately, this is often not the case for predictive policing systems, where there is evidence of bias towards age as well as race and sex leading to many people being mistakenly labeled as likely to be involved in a crime. In a system that already is under criticism for its unjust treatment of minority groups, it is crucial to find ways to mitigate this negative trend. In this work, we explored and evaluated the infusion of domain knowledge in the predictive policing system to minimize the prevailing fairness issues. The experimental results demonstrate an increase in fairness across all of the metrics for all of the protected classes bringing more trust into the predictive policing system by reducing the unfair policing of people.