Open access
Author
Date
2020-09Type
- Doctoral Thesis
ETH Bibliography
yes
Altmetrics
Abstract
Crime exacts high financial, physical, and emotional costs from individuals and societies. Therefore, there is a great deal of interest in reducing it. Criminology and other disciplines have devoted efforts to developing tools supporting crime reduction and prevention strategies. In the same vein, researchers and practitioners have shown that it is possible to curtail criminality by reducing the availability of crime opportunities. This approach has proven more effective than other measures such as, for example, imposing harsh sentences on convicted criminals with the aim of setting an example that discourages other potential lawbreakers from committing crimes.
Crime is understood as an interaction between an offender and their environment, and criminal offenders are seen as rather rational decision-makers acting upon available opportunities to maximize their own gains. Building upon these ideas, numerous strategies have emerged to identify locations that present a greater number of opportunities for criminal behavior. For these strategies to be effective, they need to be targeted at areas with a higher risk of crime. These locations are often characterized by specific features, which can be captured by data. Thus, modern approaches model crime by integrating features of the environment with historic crime data. This leads to rather accurate crime prediction models, allowing practitioners to reduce crime by planning interventions at the right times and locations.
Crime prediction models have long been built using statistical techniques. The potential of these techniques is that they can integrate a large variety of spatio-temporal features to build accurate predictive models. On the downside, they cannot account for interactions between the elements of the model on an individual level, which are very characteristic of social phenomena. Simulation techniques can overcome this shortcoming. Indeed, agent-based modeling allows one to represent dynamic individual agent behavior in its context, including interactions between the elements of the model. Moreover, these techniques allow such models to be used for testing the effect of prevention strategies on the development of crime. Existing research has explored agent-based models for theory testing, investigating the effects of prevention strategies and forecasting the development of crime. These models are mainly theory-based and use little data to instantiate a realistic environment, although some researchers call for realistic agent-based models to be able to accurately simulate crime patterns. Moreover, there is a stream of data-driven, agent-based models in other domains, which shows the successful potential of building realistic, agent-based models in a data-driven manner. With this in mind, this thesis explores how to use large-scale, static and dynamic open data to build a data-driven agent-based model in order to predict crime at a micro-level.
The overall aim of the thesis is to build a data-driven agent-based model to predict crime patterns using large-scale, openly available data sources to represent static and dynamic environmental features at street segment level. The performance of the generated patterns is tested against known crime patterns. To study the topic in more detail, this overall aim is split into two research topics.
The thesis first explores data-driven offender mobility, creating a data-driven micro-simulation generating offender mobility patterns derived from large-scale human activity data. Existing spatial agent-based models for crime instantiate mobile offenders generating crime patterns. Such offenders move along geographic units of (a representation of) the environment. Instantiating the mobility of the offenders is challenging because there is a lack of detailed information about how offenders move. Partly for this reason, the existing models have largely neglected the question of how to best implement offender mobility, focusing instead on how to represent the decision of whether to commit crimes (along the traveled paths). As a result, existing models either represent offender mobility in the form of predefined behavior (e.g. agents move between a set of predefined, random activity nodes), or they instantiate known offenders visiting known activity nodes. At the same time, an increasing amount of research is using openly available data sources to study and model human activities and mobility patterns. These data sources are used to represent mobility patterns of the general population, while offenders are said to move using the same or very similar patterns. Thus, the aim of the simulation is to generate mobility patterns that cumulatively match existing crime patterns. This shows the potential of informing offender mobility strategies using a proxy for human activity derived from location-based social networks in combination with a proxy for population flows derived from taxi trip data. This strategy is especially relevant for simulating the mobility of virtual robbers.
Moreover, this thesis uses agent-based modeling to explore offender decisions to commit crimes in a theory-informed, data-driven manner (corresponding to the second study). Theories in criminology and empirical research have long identified local, social, and environmental features as important drivers of certain criminal occurrences. Therefore, the past decade has seen the emergence of simulations (e.g. agent-based modeling) to study crime with the benefit of integrated data (with an emphasis on geographic data). The potential for such data sources to generate crime patterns similar to real patterns has already been identified. Still, data-rich crime simulations are scarce (especially data-driven simulations). In contrast, machine learning techniques are able to process large amounts of data. Thus, combining agent-based modeling with machine learning may allow one to instantiate a realistic environment based on large-scale data relevant to crime, and have agents process the data to effectively derive actionable insights. This thesis aims to do this by building an agent-based model that generates patterns similar to real crime patterns for an urban environment. For this purpose it uses large-scale static and dynamic features relevant to crime. The results of this study show the importance of adding spatial and temporal data along with agent interaction, especially to predict the top street segments most at risk of future robbery. For predicting a higher percentage of street segments at risk, spatial data prove to be the most relevant factor. This final model achieves predictive performance similar to existing models using different techniques (such as machine learning), while offering greater resolution and higher potential for practical application.
Finally, the resulting "full" crime prediction model has the capacity to be implemented as a crime prediction tool and/or an experimental tool to test the effect of crime reduction strategies on long- and short-term crime. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000451769Publication status
publishedExternal links
Search print copy at ETH Library
Publisher
ETH ZurichSubject
Crime prediction; Simulation; Modeling; Crime prevention technologies; agent-based simulation; Machine Learning; Artificial IntelligenceOrganisational unit
03681 - Fleisch, Elgar / Fleisch, Elgar
Related publications and datasets
Has part: https://doi.org/10.3929/ethz-b-000431739
More
Show all metadata
ETH Bibliography
yes
Altmetrics