Gray Matter: There’s a Fly in My Tweets

Researchers have been striving for generations to answer such questions, using health surveys of samples of individuals and computational studies of simulated populations. Now, however, the rise of social media and the burgeoning field of data science provide powerful tools to find high-precision, real-world answers with little cost or effort.

The millions of people posting to sites like Twitter and Facebook can be viewed as a vast organic sensor network, providing a real-time stream of data about the social, biological and physical worlds. While people use social media to build and maintain their social ties, the “data exhaust” of their postings can be analyzed to provide an enormous range of information at a population scale.

For example, my research group at the University of Rochester has analyzed Twitter postings from millions of cellphone users in New York City to develop a system to monitor food-poisoning outbreaks at restaurants.

We began by creating algorithms that can identify tweets about a given topic with near-perfect precision, even if the words and phrases used vary widely. The GPS information embedded in tweets sent from cellphones lets us integrate them with a variety of geographic databases.

We then feed the information into what we call the nEmesis system, whose development was led by our graduate student Adam Sadilek, now a researcher at Google. It begins by finding tweets that are sent from restaurants, which we can locate on Google Maps with 97 percent accuracy, thanks to GPS coordinates.

When a user is identified as having been at a restaurant, all of his or her tweets, from anywhere, are collected for the next 72 hours and analyzed to discover if any appear to report food poisoning symptoms, like vomiting, diarrhea, abdominal pain, fever or chills.

Such reports are rare but significant. Over a four-month period, our system collected 3.8 million tweets, from which we were able to trace 23,000 restaurant visitors and found 480 reports of likely food poisoning. Restaurants were then scored by the number of food poisoning reports from their patrons.

The Twitter reports are not an exact indicator — any individual case could well be caused by factors unrelated to the restaurant meal. But in aggregate the numbers are revealing. Working with Vincent Silenzio, who teaches in the department of community and preventive medicine at our medical school, we compared the results with the current database of restaurant inspections conducted by New York City’s Department of Health and Mental Hygiene. We found significant correlation between restaurants’ violation scores and the Twitter-based scores.

Our project isn’t alone. While an army of corporations are busy data mining social media for marketing, a small but growing number of research groups have initiated similar efforts to leverage the torrent of online information for social good.

Groups at Brigham Young University and the University of Iowa have done extensive work on influenza monitoring via Twitter posts. Researchers at Microsoft are helping to identify women who are at risk of severe postpartum depression by analyzing changes in their online behavior. And researchers at Cornell are mining the social media stream to gather data for urban planning and environmental conservation.

THE most daunting challenges in making sense of social media are data incompleteness and noise: not knowing whether you have all the information, and how to sort out what’s relevant. These problems drive fundamental research on statistical machine learning and data-mining algorithms. nEmesis and its kin provide large-scale test-beds for developing and testing solutions to these challenges.

Further, nEmesis has immediate public-policy applications. While city health inspections capture a wide variety of data that is difficult to obtain from online social media (like the presence of rodents in a restaurant’s storage room), the Twitter signal measures a perhaps more useful quantity: a probability estimate of your becoming ill if you visit a particular restaurant.

Put differently, inspections are thorough but largely sporadic. A cook who occasionally comes to work sick and infects customers for a few days at a time is unlikely to be detected by current methods. Similarly, a batch of potentially dangerous beef delivered by a truck with a faulty refrigeration system could be an outlier, but nonetheless cause loss of life.

Obviously, public-health officials can’t rely solely on tools like nEmesis. But social-media-based systems have the potential to greatly complement traditional data-collection methods, producing a more comprehensive — and timely — model for public health policy.

Henry Kautz is the chairman of the computer science department at the Hajim School of Engineering and Applied Sciences at the University of Rochester.

Gray Matter: A Focus on Distraction

TECHNOLOGY has given us many gifts, among them dozens of new ways to grab our attention. It’s hard to talk to a friend without your phone buzzing at least once. Odds are high you will check your Twitter feed or Facebook wall while reading this article. Just try to type a memo at work without having an e-mail pop up that ruins your train of thought.

But what constitutes distraction? Does the mere possibility that a phone call or e-mail will soon arrive drain your brain power? And does distraction matter — do interruptions make us dumber? Quite a bit, according to new research by Carnegie Mellon University’s Human-Computer Interaction Lab.

There’s a lot of debate among brain researchers about the impact of gadgets on our brains. Most discussion has focused on the deleterious effect of multitasking. Early results show what most of us know implicitly: if you do two things at once, both efforts suffer.

In fact, multitasking is a misnomer. In most situations, the person juggling e-mail, text messaging, Facebook and a meeting is really doing something called “rapid toggling between tasks,” and is engaged in constant context switching.

As economics students know, switching involves costs. But how much? When a consumer switches banks, or a company switches suppliers, it’s relatively easy to count the added expense of the hassle of change. When your brain is switching tasks, the cost is harder to quantify.

There have been a few efforts to do so: Gloria Mark of the University of California, Irvine, found that a typical office worker gets only 11 minutes between each interruption, while it takes an average of 25 minutes to return to the original task after an interruption. But there has been scant research on the quality of work done during these periods of rapid toggling.

We decided to investigate further, and asked Alessandro Acquisti, a professor of information technology, and the psychologist Eyal Peer at Carnegie Mellon to design an experiment to measure the brain power lost when someone is interrupted.

To simulate the pull of an expected cellphone call or e-mail, we had subjects sit in a lab and perform a standard cognitive skill test. In the experiment, 136 subjects were asked to read a short passage and answer questions about it. There were three groups of subjects; one merely completed the test. The other two were told they “might be contacted for further instructions” at any moment via instant message.

During an initial test, the second and third groups were interrupted twice. Then a second test was administered, but this time, only the second group was interrupted. The third group awaited an interruption that never came. Let’s call the three groups Control, Interrupted and On High Alert.

We expected the Interrupted group to make some mistakes, but the results were truly dismal, especially for those who think of themselves as multitaskers: during this first test, both interrupted groups answered correctly 20 percent less often than members of the control group.

In other words, the distraction of an interruption, combined with the brain drain of preparing for that interruption, made our test takers 20 percent dumber. That’s enough to turn a B-minus student (80 percent) into a failure (62 percent).

But in Part 2 of the experiment, the results were not as bleak. This time, part of the group was told they would be interrupted again, but they were actually left alone to focus on the questions.

Again, the Interrupted group underperformed the control group, but this time they closed the gap significantly, to a respectable 14 percent. Dr. Peer said this suggested that people who experience an interruption, and expect another, can learn to improve how they deal with it.

But among the On High Alert group, there was a twist. Those who were warned of an interruption that never came improved by a whopping 43 percent, and even outperformed the control test takers who were left alone. This unexpected, counterintuitive finding requires further research, but Dr. Peer thinks there’s a simple explanation: participants learned from their experience, and their brains adapted.

Somehow, it seems, they marshaled extra brain power to steel themselves against interruption, or perhaps the potential for interruptions served as a kind of deadline that helped them focus even better.

Clifford Nass, a Stanford sociologist who conducted some of the first tests on multitasking, has said that those who can’t resist the lure of doing two things at once are “suckers for irrelevancy.” There is some evidence that we’re not just suckers for that new text message, or addicted to it; it’s actually robbing us of brain power, too. Tweet about this at your own risk.

What the Carnegie Mellon study shows, however, is that it is possible to train yourself for distractions, even if you don’t know when they’ll hit.

Bob Sullivan, a journalist at NBC News, and Hugh Thompson, a computer scientist and entrepreneur, are the authors of “The Plateau Effect: Getting From Stuck to Success.”