Predicting Crime
© ThinkstockNew software uses records of previous crimes to predict areas or "hot spots" where police are then dispatched.
The field of "predictive policing" is becoming more and more common as law enforcement officials take advantage of new tools of computer science, machine learning and big data to figure out where criminals may strike next.

These programs are a far cry from Minority Report, the Tom Cruise film/Philip K. Dick novel in which citizens were arrested days or weeks before they committed crimes. But prediction methods are getting better, focusing not on an individual's brain or personality, but rather individual kinds of behavior of large groups of people -- in this case, the habits of bad guys.

Predictive policing "is not about replacing police officers with Robo-Cop," said Jeff Brantingham, an anthropologist at the University of California, Los Angeles, who has developed predictive policing software for several big city departments.

"It's about predicting where and when crime might occur."

Brantingham's software uses records of previous crimes -- their location, time of day and type of crime -- to predict areas or "hot spots" where similar events may occur. Police are then dispatched to the area to keep a lookout, or just disrupt any possible criminal behavior.

The PredPol (Predictive Policing) software program is deployed in Los Angeles, Atlanta and Tacoma, Wa., among other cities.

Officials with the Cambridge (Mass.) Police Department are working with statistics experts from the Massachusetts Institution of Technology in another direction -- trying to find patterns of behavior in the "modus operandi" previous criminal cases to stop future ones.

"We don't consider time and space, but all possible methods of the criminal," said Janet Rudin, MIT associate professor of statistics. "How did they get in? Did they push in a window or unlock a door? Did they ransack the place or leave it neatly? Do they come in while people are living there? We look at very detailed information about crimes that you wouldn't be able to look at with just hot spot analysis. It's a much harder problem."

Rudin has also helped NASCAR drivers change their tires more efficiently and predict manhole cover explosions in New York City. She says that the goal isn't to predict crime, but identify people who are committing the same kinds of crimes over and over. The algorithm especially works well for property crimes, such as burglaries and pickpockets.

In Philadelphia, scientists are joining police to hopefully curb domestic violence. University of Pennsylvania's Richard Berk, professor of criminology and statistics, is working on a new project to collect injury data from victims.

The idea is to see if certain behaviors by domestic abusers can be used to predict future violence. The program collects information about the kind of weapon used; is the offender armed with handgun or a rifle? Is there property damage? Are pets injured (a good predictor)? Are there threats of doing serious bodily harm?

"You look at location of injuries on a person's body, whether they have been strangled, that is different kind of violence than punching someone in the nose," Berk said. "When a cop walks in the door, they have to make decision whether to arrest or the individual or let him go. That decision should be affected by whether there will be violence again soon. Maybe fatal violence. We will know from these reports."

Other big data policing projects are trying to predict recidivism by looking at the likelihood that a parolee will commit another crime based on his or her past. All these researchers say the challenge is getting these software programs out of the academic setting and into the patrol car.

Rudin says that the more specific you get with data and its predictive value, the more customized and more work it is to adapt the computer algorithms to each community.

"If police don't keep track of this data," Rudin said. "We can't do it."

Policing with data is an improvement on policing without data, but it's not a replacement for the skills and training of the police officer, according to UCLA's Brantingham.

"It's not as if the data is going to do the interaction with community members," he said. "The data helps you take what are limited resources and puts them in the right place at the right time."