Quantifying Fairness in Spatial Predictive Policing

Diego Hernández, Cristian Pulido, Francisco Gómez

Submitted to Artificial Intelligence and LawMajor revisions Spotlight

Predictive policing leverages data-driven models to anticipate future criminal events and guide law enforcement strategies. However, concerns about algorithmic fairness have emerged, as these models risk perpetuating discrimination and inequities, particularly among vulnerable populations. While prior research has acknowledged the influence of disparities in crime reporting levels on these models, the extent of their impact on vulnerable populations remains insufficiently understood, posing a critical challenge in societies marked by high disparities. This study seeks to quantify the fairness of three prevalent density probability estimation models used in predictive security. Specifically, it examines their capacity to distribute benefits impartially across diverse populations in various spatial contexts. Real-world theft data were employed to calibrate three distinct predictive models, followed by the quantification of model fairness through two different measurement approaches. These measurements assess disparities in the granting of model benefits between two geographical areas-one focusing on the average prediction error benefit and the other on the utilization of the model for resource allocation. Results suggest that the predictive security models studied may be fair for the prediction but unfair over the use of the model for patrol allocation, with a maximum difference between the means of groups of 45% and an average of these differences of 12%. This highlights the nuanced nature of fairness considerations within predictive policing frameworks.

[Code] [Demo]