Predictive policing gets personal
Data mining can predict who will reoffend, not just where and when the crimes will occur.
The consolidation of data in a common data warehouse such as N-DEx is just the first step in improving nationwide investigations. The next steps involve the use of data mining techniques to predict where and when crime will occur. This sounds like a page from the script of the movie Minority Report, but the objective isn't to arrest people but to anticipate and remove the opportunity to commit crimes before they can occur.
"9-11 triggered a major paradigm shift in the policing world. It had been a response mission. Now it's a prevention mission," says Steve Ambrosini, executive director of the IJIS Institute, an industry consortium of vendors of law enforcement software.
"Predictive policing is at the top of a lot of people's lists," says Commander Scott Edson of the Los Angeles Police Department, which uses a commercial software program, Coplink, for its incident management database. "It's something that Coplink and N-DEx will mature into."
Initiatives are already well underway at the local level in areas such as Los Angeles, where the PredPol algorithm developed by UCLA has been used to analyze seven years of incident information to predict where, within 500 foot by 500 foot areas, or "predictive boxes," certain types of property-related crimes are most likely to occur during an upcoming patrol shift. And then the department can concentrate on those areas during their shifts, or can redeploy police person-power appropriately.
We're not just looking for crime. We're looking for people. Police Chief Rodney Moore, Charlotte-Mecklenburg, S.C.
In a 2012 pilot phase, the Los Angeles Police Department Foothills area recorded a 25% decrease in burglaries in a six-month period over the previous year, according to Sean Malinowski, commanding officer of the Real-time Analysis and Critical Response division. "The idea is to prevent crime."
Mathematicians developed the model based on an algorithm that's used to predict earthquakes, says Jeffrey Brantingham, chief of research and development at PredPol, a startup formed to market software that uses the PredPol algorithm to make predictions about where crimes are most likely to occur during a patrol shift to other police departments.
Like earthquakes, crime occurs along certain fault lines, and events tend to cluster together into a predictable pattern. Work now is expanding beyond car theft, burglary, and burglary from a vehicle, which represent 65% of all crimes, to predict other crimes that fall under the FBI's Part 1 classification, including robbery, rape, assaults and homicide.
The Seattle Police Department, another early user of PredPol, started out with property crime predictions, which rolled out to all five of its precincts in May. A planned July rollout of a citywide predictive model to anticipate where and when gun violence is most likely to occur has been postponed due to "conflicting priorities within the agencies," says Sergeant Christi Robbin. "Our goal is to still roll out gun violence but we do not have a set date," Robbin says.
The department hopes to stem the tide of violent crime, which has been increasing in recent years, Robbin says. But the jury is still out as to how effective it will be: "With gun crimes you have fewer incidents, so the predictions aren't as strong," she says. Within the predictions, she explains, the first three to five boxes are usually where experienced officers would expect trouble. But it's boxes 8, 9 and 10 on down the list that they never would have anticipated, she says.
Since implementing a similar predictive policing system four years ago, the City of Richmond, Va. has seen a significant reduction in all violent crimes and property-based crimes. The results were so good, in fact, that police chief Rodney Monroe implemented a similar system soon after taking on his current position as chief of police in Charlotte-Mecklenburg, N.C. That system, developed by the commercial satellite imaging company DigitalGlobe, includes historical data and refreshes every two hours to adjust predictions for 39 response areas. "We've had a 20% reduction in violent crime and a 30% reduction in property crime," he says.
Dr. Colleen McCue, senior director of social science and quantitative methods at DigitalGlobe, has been modeling violent crime using machine learning for more than 20 years. "People are creatures of habit. That's what this all goes back to," she says, adding that models can also make predictions about specific offenders if there's enough previous criminal activity.
For example, a shooter in Northern Virginia a few years ago targeted government facilities and, with the Marine Corps Marathon coming up, authorities were anxious to anticipate his next move. McCue examined previous incidents involving the shooter and ran data to create a predictive model. They discovered that the shooter preferred a position 200 meters back from the target with close proximity to a highway or a major roadway.
Using 3-D spatial data she created a heat map showing all locations on the marathon route that met the criteria. Authorities positioned people there -- and nothing happened. While it's hard to know for sure if these actions thwarted the shooter, place preference came into play once again after it became clear that this shooter also had an affinity for cemeteries. Six months later he was apprehended at Arlington National Cemetery.
See related stories:
- "It's criminal: Why data sharing lags among law enforcement agencies"
- "Cool cop tech: 5 new technologies helping police fight crime"
Predictive policing is a helpful tool, but you still need an analyst to interpret the data, rather than just depending on the system to push out all the answers, she adds. "Statistical-based approaches work in some cases, but in others [human] judgment still works better."
Charlotte-Mecklenburg, N.C., is now going beyond predicting where and when crime will occur to predict who is likely to reoffend. Instead of studying just crimes and locations to decide where crimes will occur, police departments make predictions using criminal histories to predict who will commit a crime. This approach -- making predictions about people with criminal records -- is one that both Los Angeles and Seattle have avoided due to public fears that the technology would be used to profile people based on race or the neighborhood in which they live.
In Los Angeles the program met initial resistance due to such fears. "There were some questions about whether we were violating civil rights by doing this," says Malinowski. "But we're not factoring in arrests, and there is no information about individuals. It's about crimes and the times and places they occurred."
Andrew G. Ferguson, associate professor of law at the University of the District of Columbia, has studied and published a paper on predictive policing. He says predictive policing efforts in Los Angeles and Seattle do bring up concerns about racial and class profiling, but indirectly, because it's the area, not the individual, being profiled.
"The key to determining whether predictive policing will have a discriminatory impact is to figure out if the areas targeted are disproportionality found in communities of color," he says. But, he adds, "I have heard of no major complaints from the LA rollout of the technology."
But Ferguson calls Charlotte-Mecklenburg police chief Monroe's approach of focusing on people rather than geography "troubling."
Monroe argues that the approach does not constitute profiling because the model only looks at people who already have a record of criminal activity. "We could name our top 300 offenders," he says. "So we will focus on those individuals, the persons responsible for the criminal activity, regardless of who they are or where they live."
Knowing who the bad guys are and keeping an eye on them is in itself a form of predictive policing, Ferguson says, but using computers to make predictions about one person's future behavior is a different matter. "I don't think we have the technology to know with any degree of confidence who will commit the next crime," he says.
But Monroe argues that predicting when and where known criminals offend next will be more effective than the areas-based approach taken by other agencies. What is the probability that the offender will offend again, in what timeframe and where? "We're not just looking for crime. We're looking for people," he says.
Read more about government it in Computerworld's Government IT Topic Center.