Authors: Alina Ristea*, Department of Geoinformatics Z_GIS, University of Salzburg, Austria, Bernd Resch, Department of Geoinformatics Z_GIS, University of Salzburg, Austria; Harvard University, Cambridge, Massachusetts, USA, Michael Leitner, Department of Geography and Anthropology, Louisiana State University, Baton Rouge, USA; Department of Geoinformatics Z_GIS, University of Salzburg, Austria
Topics: Geographic Information Science and Systems, Spatial Analysis & Modeling
Keywords: crime, evaluation, hotspot, prediction
Session Type: Paper
Start / End Time: 8:00 AM / 9:40 AM
Room: Wilson B, Marriott, Mezzanine Level
Presentation File: No File Uploaded
Space-time models may be used for retrospective analysis and forecasting. Predictions generated by space-time models aim to identify regions of high intensity of a specific phenomena, such as an area with high criminal activity. Several evaluation methods have been applied for space-time crime prediction processes. Whilst many types of forecasting problems exist, such as for trajectories, time series, or locations, in this analysis we focus on hotspots, highly used as output in criminology.
In this analysis, we review evaluation methods for space-time models from fields such as spatial sciences, environmental criminology, and computer science. We compare evaluation measures and their effectiveness in case studies for space-time crime prediction. Among others, evaluation matrices that will be discussed include Hit Rate (HR), Prediction Accuracy Index (PAI), Prediction Efficiency Index (PEI*), surveillance plots and confusion matrix. While several methods are frequently used for spatial crime forecasting, they have different qualities and limitations. In addition, computing accuracy scores the predicted performance can vary significantly from one time window to the other. Moreover, the spatial unit of analysis is an important component that can make a difference in the forecast.
Whilst discussing evaluation methods interpretation and selection of crime data, we will also give insight into the difference between prediction models. This study design is widely applicable in other space-time predictions from other fields.