Auditing the fairness of place-based crime prediction models implemented with deep learning approaches

Abstract

Place-based crime prediction models implemented with deep learning leverage the spatio-temporal patterns of historical crimes, together with built-environment factors, to predict aggregate volumes of crime incidents at specific locations. Nevertheless, related work has shown that crime incident data can suffer from under-reporting bias whereby certain social groups report more crimes than others and whereby certain crimes are recorded more often by the police. In addition, under the umbrella of crime opportunity theories that suggest that human mobility can play a role in crime generation, increasing attention has been paid to the predictive power of human mobility in place-based crime models implemented with deep learning. However, prior work has shown that human mobility data collected from cell phones can suffer from sampling bias because certain population groups are not equally represented due to lack of cell phone ownership. Despite these well reported data bias concerns, no related work has yet looked into the fairness of place-based crime prediction models implemented with deep learning; nor into the effect that adding human mobility data from cell phones as a predictor might have on their fairness. In this paper, we use publicly available fine-grained crime incident and human mobility data from the US to conduct a comprehensive fairness audit across multiple cities with diverse demographic characteristics, different types of crimes and various deep learning models; and we explore potential root causes of the models' fairness metrics by looking into crime data bias, mobility data bias and algorithmic bias. Our results show that unfair predictions are pervasive across place-based crime prediction models implemented with deep-learning due to the inherent data bias in the crime incident data; that adding mobility features extracted from cell phone data decreases fairness under certain circumstances; and that part of the fairness loss can be explained by bias in the crime data, and in the predictive algorithms that might be exacerbating the data bias, with the mobility data bias having an insignificant effect on fairness loss.