Hot Spot mapping uses spatial autocorrelations to identify criminal behavior in a city or other region. Traditional mapping plots the frequency of events and location and uses this to extrapolate probability trends. That does not quite work well, because crime spreads like forest fires. Criminals adapt and change their behavior. Hot spot mapping takes this into account.
The basic idea is that individuals do not make their decisions independently of one another. They act according to changing information in their environment, so their behavior is interrelated. As an analogy, humans are like ants in an ant colony, interacting, exchanging information, and displaying “swarm” behavior.
Social interactions can be considered a form of memetic diffusion. Ideas and information are transferred from one person to another. Information is “contagious” like a disease or a forest fire seeking fuel.
So agents act in local areas based on local information, but are also able to “share” information with each other so they have a macro-view of the region.
Criminals and insurgents, for instance, hear of other activity from the newsmedia. They update their behavior to reduce the likelyhood of getting caught in certain neighborhoods vs the utility of crime. This is their perceived probability of success and failure – their rationality is bounded and they are also confused by their own irrational beliefs. Some believe gambler’s fallacies for instance.
Economic-based criminal activity is the easiest to predict. Criminals are businessmen of a sort, so their behavior is very similar to normal businessmen. This is particularly true of smugglers. Other economic crimes fit this theory too, because cities have a greater economy of scale. Basically, criminals thrive in cities for the same reason corporations and factories thrive in cities. More importantly, they thrive in certain neighborhoods.
Traditional criminal mapping is simple. You put a pin on the map everywhere there is a crime and do simple pattern recognition. You will see clusters in one neighborhood and almost nothing in others. More advanced probability statistics measured the frequency of each crime across space and time and “projected” this frequency into the future – which works reasonably well, but obviously is not as predictive as we need it.
Spatial Autocorrelation goes one step further than traditional mapping. It takes into account the “contagion” like behavior of information and criminal responses. So far as I know, this originated with the epidemiology statistics.
Spatial autocorrelation is a form of multivariate analysis. First, what it does is correlate the variable with the location of the variable. Then you assess whether or not the variables are correlated with location. This confirms the spatial pattern. Second, you determine if multiple variable interact with one anther or if they are independent.
So for crime, insurgency, and other violent activities, we know there is informal and formal interactions. Spatial autocorrelation reveals the process driving the patterns. It allows superior prediction to traditional frequency measurements because it takes into account adaptive behavior.
Then you move on to the meat of this data mining by mapping out criminal behavior. This produces statistical data like “journey-to-crime” mapping which shows which criminal travel routes and how far from their home base they will operate.
It is a cliché to say that crime spreads like a disease, but previous work by Dr Bowers and her colleagues found that this is exactly how crime does spread. Using statistical techniques developed to study the transmission of infections, they found that burglaries cluster in space and time in predictable ways. For example, properties within 400 metres of a burgled home, particularly those on the same side of the road, are at an increased risk of being broken into for up to two months after the initial incident.
Using these and other findings, the team created algorithms that predict where criminals will strike next, and then used those algorithms to generate “prospective hot-spot maps”. These divide an area into 50-metre squares—a level of resolution chosen because 50 metres is a typical line-of-sight for a police officer in an urban area—and give a crime forecast for each square.
In their paper, Dr Bowers and her colleagues reveal the results of a study of burglaries in Merseyside, in northern England. Using historical data, they pitted their predictive modelling method against two traditional crime-mapping systems. They found that their method successfully “hindcasted” 62-80% of burglaries. The traditional techniques, by contrast, hindcasted only 46% of those incidents.
The newest anti-crime datamining, like Crimestat, use hot spot mapping.
Here’s another example. Police can identify criminal clusters and predict that a homocide cluster will emerge in this region. And so it does. Murders occur where there is a higher frequency of lesser crimes, possibly because it is related to lesser crimes like drug use and smuggling.
This will be a valuable tool for crime-fighting and urban warfare in the 21st century.
This is how nerds fight back.