Police forces around the world are turning to new technologies to help them be more effective in crime prevention with the same number of officers. One of the most interesting emerging technologies in law enforcement, predictive policing, focuses on crimes that have not yet taken place, but are likely to occur in the future.
Predictive policing uses statistical and analytical techniques to construct patterns of crime occurrences and help forecast various aspects of crime. Using historic crime reports and, in some cases, other data sources like social media, statistical algorithms can identify most probable future locations, times, and types of crimes. LEAs can use crime predictions to patrol areas where crime is anticipated at a certain time in order to increase the chances of deterring criminal activity.
While methods of predictive policing are most commonly used to model risk profiles of most probable crime locations, they can also be utilized to model risk profiles of offenders and predict how likely they are to repeat their offence or engage in other criminal activity, or even predict victims of crime.
Mapping and mathematical modelling of crime data for crime forecasting started in the 1990s, but growing computing capabilities and development of machine learning techniques in the last decade enabled analysis of large crime data sets and ever more accurate predictions. Predictive policing is being reportedly used today by LEAs in several US states, UK, Germany, the Netherlands, Switzerland, and China.
Some examples of predictive policing software solutions include: PredPol, Precobs, CrimeRadar, and HunchLab.
Probably the most famous of them is PredPol, developed based on the idea that models, similar to those describing seismic activity, can be used to model crime patterns. According to the developers from the University of California, Los Angeles, additional crimes often follow the initial event in the near time and space, comparable to a seismic aftershock. PredPol uses only three data attributes of each historic crime report: crime type, location, and time, to generate maps of areas where crime is most likely to occur for each police shift of any day.
Another predictive policing system, Precobs (Pre Crime Observation System), was developed in Germany and focuses on so called “near repeat crimes” – those that are frequently repeated within three days in a close proximity. The most typical such crimes are burglary, street robbery, armed robbery and motor vehicle theft. Precobs uses more specific data about past burglaries, such as the kind of items that were stolen and the way the burglar entered a house (a more thoughtful modus operandi would suggest a higher recurrence rate). Theoretical explanations of near repeat crimes are based on the hypotheses that past victimisation increases the probability of becoming a victim again and that offenders often return to the place of their first crime to use their knowledge about the area.
CrimeRadar, an application developed by the Brazilian Igarapé Institute in 2016, uses crime records for Rio de Janeiro since 2010 to show locations, times and types of previous crimes, and to predict future risk levels for areas of the city. CrimeRadar is especially interesting because it is publicly available – it aims to improve personal safety and security of citizens and visitors to the city by providing them with predictions of which areas are the most dangerous at certain times. A similar crime predicting map, Kapo Aargau app, is provided in Switzerland as a mobile application.
MEDIA4SEC’s Report on the State of the Art marks a need of information sharing among LEAs to improve predictive policing models and decisions based on them. The common denominator of all modern predictive policing solutions is the use of machine learning algorithms, with differences in the variety of data, used as input. To extend and improve predictions, studies have suggested to include social media activity, e.g. Twitter messages, as a source of input. Researchers from the University of Virginia claim to significantly improve crime prediction accuracy by adding linguistically analysed Twitter messages and their location information.
Another proposed extension of input data is gunshot location information from automated gunfire detection systems. An example of such system is ShotSpotter, deployed in many cities across the US. Using microphones dispersed around the city, it can detect gunshots and calculate the location where the weapon was fired.
Threats and criticism regarding the use of personal information like posts from social media in predictive policing are also worth mentioning. Critics suggest that monitoring social media can lead to discrimination, over-criminalization and unwarranted monitoring of youth, minority groups, or activists in case riots and protests’ predictions. Another concern is that the use of such data would be used as justification for expanding the collection of citizens’ personal data by the authorities. Even when no personal data is being analysed, predictive policing systems might cause law enforcement agents and citizens (especially in cases where the predictions are public, like CrimeRadar) to exaggerate dangers of certain neighbourhoods and can lead to further segregation. Potential risks of predictive policing were already discussed in MEDIA4SEC’s Ethics and Legal Issues Inventory (D1.3).
Although predictive policing techniques are already in use in many countries, there is still a lack of studies and reports about their effectiveness, while the societal effects remain debatable. As a topic between everyday policing and technological solutions, predictive policing will be discussed in the next MEDIA4SEC workshop in Barcelona as well as in our last workshop, in September 2018 in Brussels.