This proposal examines the relationship between crime, social vulnerability, and policy outcomes through interpretable machine learning, specifically using explainable boosting machines (EBMS)

The National Institute of Justice (NIJ) awarded a $180,000 Project Grant to the University of Illinois to examine the relationship between crime, social vulnerability, and policy outcomes through interpretable machine learning, specifically using Explainable Boosting Machines (EBMs). The 5-year project aims to investigate socioeconomic disparities and their influence on crime prediction in communities, which may bias algorithmic predictions in predictive policing and hotspot identification. The study will leverage crime event data, socioeconomic characteristics of communities via the Social Vulnerability Index, and spatial analysis to explore transparency and accountability issues in crime prediction algorithms. The results will be made available as an open resource to enable further research on addressing social vulnerability and crime.

Central to the investigation is the exploration of socioeconomic disparities and their influence on crime prediction in communities, which may bias algorithmic predictions. Traditional methods in predictive policing and hotspot identification often rely on opaque, black-box algorithms, making IT difficult to understand why a particular prediction is produced. This opacity and bias can have policy challenges within the criminal justice system, particularly when the decisions made by algorithms come under scrutiny. Such opacity, along with potential bias, raises concerns and the need to address transparency and accountability issues. Driven by the need to advance beyond conventional approaches, the study aims to illustrate the complex, dynamic relationship between crime and social vulnerability. The study leverages three datasets that encompass crime events at local and national levels and the socioeconomic characteristics of communities, via the social vulnerability index (SVI). A holistic analysis will examine representational and algorithmic bias in predicting crime based on social vulnerability characteristics of communities. Additionally, the proposal aims to explore how spatial effects, specifically spatial heterogeneity and spatial autocorrelation, can be interpreted and analyzed in EBMS. Understanding spatial effects can provide valuable insights for resource allocation and disparity analysis. Further examination will focus on the downstream implications of systematic bias and errors in crime prediction that arise from using social vulnerability features of communities. This work will be made available as an open resource, enabling practitioners, researchers, and the public to explore the inner workings of algorithmic predictions in the context of community disparities. This proposal provides tools and insights for further research to strive for a more informed and supportive approach to addressing the nuances of social vulnerability and crime. CA/NCF