Solutions to Reinforced Feedback Loops in Machine Learning-based Predictive Policing
No Thumbnail Available
Predictive Policing is an increasingly popular prediction system used by police all over the world to make tactical decisions on patrolling based on historical crime data of a variation of geographical locations. Results are considered to be positive; however, the proven downside of the software is reinforced feedback loops. Increased patrolling rates in a location, likely finding crime, and reporting back to the algorithm that indeed crime has been found can result in bias against certain locations and its inhabitants. Using a simplified implementation of a predictive system, countermeasures that can reduce this affect are tested and judged on their effectiveness. The successful countermeasures consist of connecting population numbers to crime rates, using precincts instead of geographical locations, and/or adding more attributes like type of crime. It is important to be aware of ways to reduce a feedback loop in these location-based systems without losing performance.
Faculteit der Sociale Wetenschappen