In the early days of McDonald’s, visionary co-founder Ray Kroc adopted an idiosyncratic approach to site selection. Dispatching the company plane, “we used to spot good locations for McDonald’s stores by flying over a community and looking for schools and church steeples,” he wrote in his autobiography, Grinding It Out.
Today, competing chains are more likely to launch drones to count the number of cars in their competitors’ parking lots, bringing machine vision and learning to bear on geospatial information. It’s one thing to count cars or steeples, though quite another to train an algorithm to determine the make and model of each vehicle, to marry its observations with demographic and geographic datasets, and then teach an AI to make its own classifications and conclusions.
“We look at AI as a way to fuse all of these variables with geospatial data into something we can use,” says Mansour Raad, a senior software architect and Big Data lead at Esri, the global market leader in geographic information system (GIS) software, location intelligence, and mapping. “Using machine learning, the system can find those patterns for itself. The sum of these things is amazing.”
Using location intelligence for site selection is nothing new, of course. Even Ray Kroc had such tools at his disposal. (“We have a computer in Oak Brook that is designed to make real estate surveys,” he wrote. “But those printouts are of no use to me.”)
What’s different this time are advances in hardware and software, and the emergence of the geospatial cloud, that have made it remarkably cheap to store, combine, and analyze large datasets, coupled with the maturation of machine-learning techniques to discover patterns humans may have missed. The combination is bringing a fresh set of eyes to solving persistently thorny real-world problems.
REDUCING HIGHWAY FATALITIES
A particularly urgent example is traffic fatalities. Last year in the United States, 37,133 people were killed in vehicle crashes, according to the National Highway Traffic Safety Administration—a slight decline from the previous year, but still a precipitous increase from earlier in the decade. Not all crashes are distributed equally—a tiny percentage of American roads account for an alarmingly high percentage of deaths and serious injuries. The trick is understanding why.
One approach is “density-based clustering.” Starting with NHTSA data on the location of each crash, Esri researchers first set out to map the largest clusters in the United States, a fairly simple task. Then they asked what each cluster might have in common with the others, enriching the data with additional variables and attributes: How close is each incident to an intersection? Are there stoplights or billboards nearby? What is the curvature of the road, if any? Which of these variables (or combinations of variables) resulted in each fatality?
Traditional techniques would analyze each crash through the lens of a particular hypothesis—maybe it’s due to an “out-of-context curve.” The advantage of machine learning, Raad says, is that rather than pre-selecting seemingly relevant variables using human intuition, it’s easier to give the machine all the variables—hundreds or even thousands—and let it sort them out.
“We just gave it simple rules and told it to bring us back what it found,” Raad says. In this case, what it found was combinations of attributes that likely played a role in each crash. These combinations could then be used by NHTSA to prioritize which clusters to tackle first.
For another traffic-related project, Esri was tasked by a stadium owner in the Southeast to help safely untangle pedestrians from cars after games. Integrating traffic signals with real-time video feeds and image detection, its engineers employed a machine-learning technique called “trend detection” to analyze footage, predict when and where crowds would form, and manage vehicles accordingly. The alternative would have been to close streets entirely, snarling traffic, or request police for crowd control. The AI-driven approach increases safety while reducing the strain on the stadium’s neighbors.
In both cases, the fusion of machine learning and location intelligence produced new opportunities to intervene—whether to straighten a dangerous curve or synchronize traffic signals to changing conditions. Insights led to action.
A NEW WAY OF THINKING ABOUT INSURANCE
Esri isn’t alone in thinking this way. In Chicago and Washington, D.C., a startup named Open Data Nation (ODN) is working on behalf of each city government to predict the risks of a collision, block by block. “In Washington, D.C., people of color are seven times more likely to be struck by a vehicle,” says Carey Anne Nadeau, ODN’s CEO. “We can use data to decide which intersections are the most dangerous, and whether to implement a speed bump or other measures by supplementing decisions with data.”
Nadeau describes her company’s mission as “measuring the risk around you,” emphasizing the role place and proximity play in the equation. For similar reasons, the insurance industry has been quick to embrace location intelligence and AI as a means for personalizing premiums to reflect individual behavior, rather than demographics.
For example, John Hancock Insurance now requires new holders of its life-insurance policies to report their eating, drinking, and exercise habits in exchange for lower premiums and financial rewards. A November study commissioned by the company found that policyholders wearing Apple Watches to capture personal data increased their average daily exercise by 34 percent.
Firms keen to implement the same approach for auto insurance have turned to Esri and others to explore coupling onboard accelerometers with real-time geospatial data and machine learning to recognize risky behaviors such as speeding and distracted driving. A series of small, quick corrections while steering through a curve, for instance, might be flagged as texting—with immediate ramifications for their policy.
“Every company I speak to is pursuing moment-by-moment premiums,” Raad says. “One wanted automakers to install a set of green, yellow, and red lights on the dash to warn them about their driving. It was a crazy idea, but you can start to imagine automakers and insurance firms working together symbiotically.”
Whether public safety or insurance, retailing or logistics, finance or manufacturing, the merger of location intelligence with ever-more-powerful machine-learning techniques promises to unlock new insights from old problems. “With some of the older statistical methods, you reach a plateau in terms of how much value you gain out of data,” explains Alberto Nieto, an AI expert and GIS solution engineer at Esri. “But in the newer deep-learning methods, the more data you throw at it, the more value you keep extracting out of that data.”
This is true in site selection as well. Drones aside, many firms still rely on what’s known as “drive-time analysis” to calculate the geographic customer base for a bank branch or pharmacy, for instance. Adding AI to the mix offers organizations a deeper understanding of what defines a desirable location, just as marrying car crash data with stoplight positions might yield an otherwise hidden relationship.
In a bank’s case, machine learning reveals a finer-grained analysis of who and where its customers are likely to be. This, in turn, offers executives the ability to rapidly model revenues projections for a single branch or network of branches. “Rather than think of a single site in isolation, we can map multiple locations that work in cooperation, then combine that with customer segmentation,” Nieto says.
It’s a far cry from Ray Kroc flying above newly built suburbs looking for schools, but the principle remains the same—location plus information plus intuition can equal predictions capable of delivering sustainable competitive advantage. The intelligence may be artificial, but the results are anything but.
This article was created for and commissioned by Esri.