Date of Award
Master of Science (MS)
We examine in detail two geospatial analysis algorithms commonly used in predictive policing. The k-means clustering algorithm is used to partition input data into k clusters, while Kernel Density Estimation algorithms convert geospatial data into a 2-dimensional probability distribution function. Both algorithms serve unique roles in predictive policing, helping to inform the allocation of limited police resources. Through critical analysis of the k-means algorithm, we found that parameter choice can greatly impact how crime in a city is clustered, which therefor impacts how mental models of crime in the city are developed. Interviews with crime analysts who regularly used k-means revealed that parameters are overwhelmingly chosen arbitrarily. Similarly, KDE parameters greatly influence the resulting PDF, which are visualized in difficult to interpret heatmaps. A mixed method user study with participants of varying backgrounds revealed that those with backgrounds in law enforcement and/or criminal justice rarely actively chose the parameters used, in part due to not fully comprehending the meaning of less obvious parameters. It was also found that individuals with different backgrounds tended to interpret heatmaps and make resource distribution decisions differently.There are several implications from these findings. Primarily, this implies that most would-be users lack the training and expertise to reliably implement and interpret geospatial crime analysis algorithms. Both within and without crime labs, critical thought is rarely given to parameter choice, especially for parameters without a clear, easily understandable explanation. These factors illuminate predictive policing being an inexact science, despite being taken as reliable and objective. These shortcomings and misconceptions, due to their pivotal role at the earliest part of the policing and criminal justice system, have long term consequences for denizens of any place being policed at behest of an algorithm.