Crime Hotspot Detection: A Computational Perspective

Introduction

Analyzing crime locations and using the spatial information associated with them is a fundamental task of environmental criminology since the main goal of a crime investigation is to locate the criminal and/or prevent more crimes from occurring. Two important spatial theories in environmental criminology are routine activity theory and crime pattern theory. Routine activity theory states that a crime location is related to a serial criminal’s frequently visited areas (Burn, 1982). Crime pattern theory extends routine activity theory on a spatial model (Brantingham & Brantingham, 1993). When crime analysts study the locations of crime sites and make inferences that help track down a serial criminal or that identify areas where extra police presence is needed, they are using these theories. Thus, the motivation of crime hotspot detection can be summarized as:

• 1.

  To detect current/emerging crime hotspots to focus the deployment of police forces, and

• 2.

  To predict the location of a serial criminal’s residence.

Previously, most analysis of crime locations was done manually. However, the task of enumerating potential hotspots by hand and detecting meaningful patterns in the data is arduous, even for experienced analysts. Rising crime numbers in larger cities increase analysts’ workload and compound their stress. For example, a typical crime dataset of the continental United States includes around 10⁷ crimes annually (FBI, 2015). Therefore, law enforcement agencies need computational tools that can automate hotspot detection. Moreover, these tools should eliminate the chance patterns (false positives). Most police departments and similar entities have limited resources for crime mitigation and prevention. Therefore, false positive results, that is, hotspots that occur only by chance, risk causing real harm. If a police department diverts money and manpower to a chance hotspot location that don’t need it, may leave other areas less protected. Chance hotspots can have other harmful consequences as well. When police increase their presence in a location identified as a crime hotspot, people naturally begin to avoid the area out of fear for their safety, and the neighborhood may become stigmatized.

Hence there is a growing need for crime hotspot analysis tools which can handle large crime datasets as well as eliminate chance patterns. Such tools may help crime analysts accomplish their law enforcement goals with optimal resource allocation (Eck & others, 2005).

Many computational techniques (Spencer Chainey & Ratcliffe, 2013; Eck & others, 2005; Levine, 2006) are used for hotspot detection. One of the most common methods is to enumerate hotspots with clustering techniques. These techniques are highly scalable and widely used in many societal applications (Ester & others, 1996; Ilango & Mohan, 2010; Liao, Liu, & Choudhary, 2004; Sarmah, Das, & Bhattacharyya, 2007; Wu et al., 2008). Nevertheless, clustering has two important limitations for hotspot detection. First, these techniques require the user to input the number of clusters beforehand. This does not conform well to hotspot detection, whose aim is to search for an abnormality that may or may not exist in the data. In other words, the user does not know beforehand whether a hotspot exists, or if so, how many. A second limitation is that most clustering techniques lack a statistical significance test to eliminate chance occurrences of hotspots.

This chapter focuses on statistically significant hotspot detection techniques that aim to remove chance patterns. The shape of the desired hotspot, determines the enumeration technique used to generate candidate hotspots and spatial scan statistics is used to assess the statistical significance of each candidate.