Designing a predictive policing system
The best models result when every member of the agency is involved in the analytical design
The latest technological golden boy, predictive policing, is producing positive results in several jurisdictions across the country. Even the old school skeptics are coming to appreciate there could be some traction to the new methods.
One of the premier vendors of this technology is IBM. Much of it is built on two acquisitions — i2 and COPLINK — of what used to be separate companies and inventions. COPLINK was an information-sharing network where diverse local agencies spread across a region could make use of one another’s selected crime and records data through an automated mobile and fixed-station network.
Since criminals don’t respect jurisdictional boundaries, it was possible for a crook living in one area to be very active in another, and not have the police in his hometown know a thing about what he was up to.
COPLINK not only allowed different agencies to share this information, but provided a platform where dissimilar databases became readable for one another over a common interface.
The other major acquisition was i2, otherwise known as “the investigator’s notebook.” i2 made sense of millions of items of disparate data such as vehicle makes, models and license plate numbers, traffic citations, telephone pen register number lists, tax assessor records of property ownership, and lots of other minutiae.
By comparing all the items with all the others, the software created complex link charts showing how every piece of information was or wasn’t related to the others.
An investigator might not know of the relationship between Person A and Person B until i2 pointed out that Person A had received a traffic ticket while driving a car belonging to someone who frequently called Person B’s phone number. No human could make these associations as thoroughly as the i2 software.
Many of IBM’s predictive policing modules or features are labeled “IBM i2 COPLINK” Face Match, Computer Statistics, Dashboard, etc., illustrating the genius of the technology.
Chris Knisley, Worldwide Business Development Manager at IBM, described some of the real-world applications of their products in a recent interview with PoliceOne.
“Tuscon was the very first COPLINK customer for us. They’ve been able to take all of their information from their records management system, from the various databases they’ve got, bring it together, normalize it, and make it available for predictive policing,” Knisley said.
Knisley pointed to a specific example of an anecdotal model where if two or more auto thefts occur in a particular area in a 24-hour period, then the likelihood of a gang drive-by in another area is significantly increased.
“So we know we want to send resources to that area where the gang violence is going to occur, based on a figure that came from a completely different time and place.”
While it’s nice to catch the actors in the gang drive-by, it’s even better if the drive-by doesn’t take place at all. Crime rates have dropped in the cities where predictive policing has been deployed.
In Santa Cruz (Calif.), burglaries dropped 27 percent in a month, compared to the same month the previous year. The Foothill Division of LAPD saw burglaries drop 33 percent and violent crime reduced by 21 percent.
In some cases, this is because the bad guys have been identified and arrested. In others, it’s because the crooks didn’t like the excess attention they were getting and didn’t bother to do the crimes at all.
Information sharing and inter-agency cooperation is obviously of paramount importance here. In the above scenario, if the car theft takes place in Agency A’s territory and Agency B with the drive-by doesn’t know what Agency A is up to, it’s unlikely any connection will be made. When these systems are put in place, there is a direct relationship between the number of partners and the systems’ effectiveness.
In building the system, the designers from both the law enforcement agency and the software company have to talk to the cops. Many cops intuitively know that an event is going to spark an individual or a string of crimes, even if they don’t know they know.
The best models result when every member of the agency is involved in the analytical design.
Some information can come from structured surveys, but just as much is likely to result from informal interviews with both rookie and veteran cops. Everyone sees the world in a slightly different way, and the union of these perspectives are what make the analysis more effective.
“There are so many officers who have that ‘art of policing’ knowledge, that intuitive knowledge formed from years on the beat,” Knisely said.
“They know the patterns, they know what to look for, but it’s all in their head. So, to me, the first step is to get that data together, normalize it and prepare it for use. You can have the expertise to build integrated models, but if you don’t have the underlying data in a useful format, you’re never going to get the models to work.”