Predicting crime with Twitter

When a single gunman, Ismaaiyl Brinsley shot and killed two policemen on December 20 in New York, he both warned pedestrians and bragged about his intentions on the photo sharing platform Instagram. Police in major cities in the US, are known to use internet tools like Google Earth or Twitter to monitor movements or alert them to certain events. The New York Police Department had also shown interest in Gerber's study. So could increased social media monitoring predict the behavior of people who are going to commit a crime in the future?

How did you come up with the idea of developing a tool to predict crime?

The research lab that I've been in since 2011, builds statistical models of crime. The idea is to take the historical record of crime in a city like Chicago and identify correlations in it. For example, the correlation of robbery locations and lighting infrastructure in different neighborhoods. What the models do is use those correlations to project where crime is most likely to occur.

And what made you use Twitter activity to predict crime?

If you think of Twitter as just an additional layer of information that blankets the city, the same idea applies. Instead of looking at road networks and lighting infrastructure, you look at the layer of tweets that are posted across the city.

So you need to be in a city where people tag their tweets with GPS coordinates. You can picture them as pins in a map with messages attached to them. So we analyze those pinned down messages and look for correlations between the occurrence of theft and what people are discussing in their tweets in the period leading up to those crimes.

For which types of crimes would you see high correlations?

The crimes that tend to work best are the frequent crimes, so assault or theft. Crimes like homicide – Chicago sees a few 100 homicides a year, that's about a tenth of the number of thefts and assaults – so when we watch these crimes, we don't see quite as a good a performance when we look at the Twitter data.

What kinds of things are you looking for? Do people actually announce ‘I'm about to break into a shop,' or do you see it through their online behavior?

They don't announce it for obvious reasons. Police departments monitor Twitter routinely. Quite often they have analysts, part of whose job it is to look at tweets in the area and monitor things. Criminals that and for that reason, people never post information about crimes they are going to commit. Gangs might be the exception. I haven't studied gangs, but there is research that looks into gangs and their use of social media. They get online and openly taunt police and talk about what they are going to do.

But even though most people don't talk about criminal things on Twitter, they do talk about other routine activities that they engage in. For example going downtown to bars or clubs. And those activities do correlate with crime. When you get a bunch of people drinking and socializing, you're bound to have problems in some cases.

Would this also work in instances like the New York police shootings, where the person was a lone gunman?

Attacks which are carried out by a single individual are a different sort of problem where the approaches currently taken might not work. This is because we rely on a large number of people talking about things ahead of time.

Has your research been used yet by the police?

Not the Twitter aspects. Lots of police departments use the non-Twitter version, where they look at the historical crime records. Chicago and NYPD are looking into it, but to my knowledge they haven't started doing anything. What it would take, would be demonstrating that when you look at the data from Twitter you can actually improve predictions enough to have positive outcomes on police response times and crime rates.

On the one hand, it seems natural for police to use Twitter, but what if it gets out of hand? What if police start targeting groups or individuals? Or if state starts cracking down on opposition members like what happened in countries like Bahrain.

These sorts of things cut both ways. In some cities that use our methods, they've shown pretty dramatic decreases in crime. Because you know where to put the police to cutback crime. That's a big plus and that's how I hope all of this will be used. But of course you run the risk of identifying factors in your analysis that pick out specific people or specific groups.

With Twitter you get closer to monitoring people and their everyday communication, because Twitter is very public. In my mind there is no sure way to prevent those sorts of abuse, other than people being aware of what Twitter does in terms of sharing data and being aware of what is going to happen to your posts when you send them out. I'm guessing a lot of people don't really know the full extent to which tweets can be shared and redistributed by anybody.

Are other social media platforms like Facebook also suited for crime prediction?

It's trickier. Facebook is pretty locked down. When you sign up with your account it's all quite private and you have to go in and share more information. Twitter is the opposite. By default it is very public and you have to do certain things, like send private messages, to make it more private.

Mathew Gerber is a researcher at the predictive Technology Laboratory of the University of Virginia.