FASE: A Fairness-Aware Spatiotemporal Event Graph Framework for Predictive Policing
Summary: arXiv:2604.18644v1 Announce Type: cross
Abstract
Predictive policing systems that allocate patrol resources based solely on predicted crime risk can unintentionally amplify racial disparities through feedback driven data bias. In response to this critical issue, we present FASE, a Fairness Aware Spatiotemporal Event Graph framework. This innovative framework integrates spatiotemporal crime prediction with fairness-constrained patrol allocation, complemented by a closed-loop deployment feedback simulator.
Model Overview
To demonstrate the efficacy of FASE, we model the city of Baltimore as a graph comprising 25 ZIP Code Tabulation Areas. Our analysis utilizes a dataset of 139,982 Part 1 crime incidents collected from 2017 to 2019, analyzed at an hourly resolution. This data generates a sparse feature tensor that serves as the foundation for our predictive modeling.
Prediction Module
The prediction module of FASE employs a combination of a spatiotemporal graph neural network and a multivariate Hawkes process. This dual approach is designed to effectively capture spatial dependencies and self-exciting temporal dynamics associated with crime incidents. The outputs from this module are modeled using a Zero Inflated Negative Binomial distribution, which is particularly suited for handling overdispersed and zero-heavy crime counts. Our model achieved a validation loss of 0.4800 and a test loss of 0.4857, indicating its predictive accuracy.
Fairness-Constrained Patrol Allocation
Patrol allocation within FASE is framed as a fairness-constrained linear optimization problem. The objective is to maximize risk-weighted coverage while adhering to a Demographic Impact Ratio constraint, which is permitted a deviation of no more than 0.05. Over the course of six simulated deployment cycles, the fairness metrics remained consistent, falling within the range of 0.9928 to 1.0262. Furthermore, the coverage achieved during these cycles ranged from 0.876 to 0.936.
Results and Implications
Despite these advancements, our findings reveal a persistent detection rate gap of approximately 3.5 percentage points between minority and non-minority areas. This gap underscores a critical insight: while allocation-level fairness constraints can help mitigate bias, they are insufficient on their own to eradicate feedback-induced bias in retraining datasets. This highlights the pressing need for comprehensive fairness interventions throughout the entire predictive policing pipeline.
Conclusion
The FASE framework represents a significant step forward in the pursuit of fairer predictive policing practices. By integrating fairness considerations into spatiotemporal crime prediction and patrol allocation, FASE aims to reduce racial disparities and improve the overall effectiveness of policing strategies. Ongoing research in this field is essential to address the complexities of bias in law enforcement practices and to develop robust solutions that prioritize equity.