Bayesian models in geographic profiling

TL;DR

This paper introduces a Bayesian approach to geographic profiling that classifies offenders into behavioral types and improves anchor point estimation, especially for non-residents, outperforming previous methods like Rossmo's.

Contribution

The paper develops a Bayesian model that incorporates offender classification, directionality, and residency status, enhancing geographic profiling accuracy over existing methods.

Findings

Improved accuracy in locating offenders' anchor points.

Effective classification of offenders into behavioral groups.

Significant performance gains for non-resident offenders.

Abstract

We consider the problem of geographic profiling and offer an approach to choosing a suitable model for each offender. Based on the analysis of the examined dataset, we divide offenders into several types with similar behavior. According to the spatial distribution of the offender's crime sites, each new criminal is assigned to the corresponding group. Then we choose an appropriate model for the offender and using Bayesian methods we determine the posterior distribution for the criminal's anchor point. Our models include directionality, similar to models of Mohler and Short (2012). Our approach also provides a way to incorporate two possible situations into the model - when the criminal is a resident or a non-resident. We test this methodology on a real data set of offenders from Baltimore County and compare the results with Rossmo's approach. Our approach leads to substantial improvement over Rossmo's method, especially in the presence of non-residents.