Addressing Under-Reporting to Enhance Fairness and Accuracy in Mobility-based Crime Prediction

Abstract

Traditionally, historical crimes and socioeconomic data have been used to understand crime in cities and to build crime prediction models. Nevertheless, the increasing availability of mobility data from cell phones to location-based services, has introduced a new family of mobility-based crime prediction models that exploit the relation between mobility patterns and reported crime incidents. One of the major concerns of using reported crime data is underreporting, which will bias the crime predictions. In this paper, we propose a novel Bayesian Hierarchical model that utilizes domain knowledge about biases in reported crime data to characterize and enhance fairness and accuracy in mobility-based crime predictions. An in-depth feature analysis reveals the influence that various factors might play in crime under-reporting and algorithmic fairness for mobility-based crime predictors.