With data from 42 police precincts on crimes committed between January 2010 to March 2016, CrimeRadar tracks some 14 million different crime events. But the app goes beyond mapping historical crimes: Through machine learning and predictive analysis, CrimeRadar will also map out future crime trends—like an open-gov pre-crime heat map…
Muggah says that Igarapé struck a deal with the Institute for Public Security, a state government agency, to build a public-facing mobile app that would show the distribution, intensity, and typologies of crimes across metro Rio. The researchers analyzed data centralized with the ISP along with data from Rio’s 190 system (like 911 in the U.S.) and created 812 categories for crimes. Those break down into capital crimes and violent crimes (like armed assault or intentional homicide), less-intense crimes (thefts, burglaries), and “victimless” crimes (loitering, prostitution).
“We built out a model that uses three data points—the time, the location, and the event—by discriminating in geospatial polygons using these three tiers,” Muggah says. “This algorithm creates a score, a risk score, based on those three data points, for every 250-meter-by-250-meter square unit in the state. You group some of the hundreds of thousands of scores for each sector into deciles to create a simplified, color-coded risk rating, on a scale of 1 to 10.”…
“We have over an 85 percent accuracy of mirroring risk against actual events. The beauty of machine learning is that this improves over time,” Muggah says. “The more data, the more information you feed into it, the higher-resolution your risk projections are going to be.”
Two things strike me as interesting:
- The claim that this is for the good of individuals who will be able to then make decisions. What about promoting the public good? This reminds me of apps in the United States that identified tougher neighborhoods but then received backlash.
- I’m not sure that 85% accuracy is good or bad. Obviously, such models strive to be much better than that. At the same time, making predictions (and with increasing levels of accuracy regarding times, locations, and actors) in a large city with many variable factors (particularly humans) is difficult. It will be interesting to see how accurate these models can be.