Via MIT Technology Review
-----
Researchers have created software that predicts when and where disease outbreaks might occur based on two decades of New York Times articles and other online data. The research comes from Microsoft and the Technion-Israel Institute of Technology.
The system could someday help aid organizations and others be more
proactive in tackling disease outbreaks or other problems, says Eric Horvitz,
distinguished scientist and codirector at Microsoft Research. “I truly
view this as a foreshadowing of what’s to come,” he says. “Eventually
this kind of work will start to have an influence on how things go for
people.” Horvitz did the research in collaboration with Kira Radinsky, a PhD researcher at the Technion-Israel Institute.
The system provides striking results when tested on historical data.
For example, reports of droughts in Angola in 2006 triggered a warning
about possible cholera outbreaks in the country, because previous events
had taught the system that cholera outbreaks were more likely in years
following droughts. A second warning about cholera in Angola was
triggered by news reports of large storms in Africa in early 2007; less
than a week later, reports appeared that cholera had become established.
In similar tests involving forecasts of disease, violence, and a
significant numbers of deaths, the system’s warnings were correct
between 70 to 90 percent of the time.
Horvitz says the performance is good enough to suggest that a more
refined version could be used in real settings, to assist experts at,
for example, government aid agencies involved in planning humanitarian
response and readiness. “We’ve done some reaching out and plan to do
some follow-up work with such people,” says Horvitz.
The system was built using 22 years of New York Times archives, from 1986 to 2007, but it also draws on data from the Web to learn about what leads up to major news events.
“One source we found useful was DBpedia,
which is a structured form of the information inside Wikipedia
constructed using crowdsourcing,” says Radinsky. “We can understand, or
see, the location of the places in the news articles, how much money
people earn there, and even information about politics.” Other sources
included WordNet, which helps software understand the meaning of words, and OpenCyc, a database of common knowledge.
All this information provides valuable context that’s not available
in news article, and which is necessary to figure out general rules for
what events precede others. For example, the system could infer
connections between events in Rwandan and Angolan cities based on the
fact that they are both in Africa, have similar GDPs, and other factors.
That approach led the software to conclude that, in predicting cholera
outbreaks, it should consider a country or city’s location, proportion
of land covered by water, population density, GDP, and whether there had
been a drought the year before.
Horvitz and Radinsky are not the first to consider using online news
and other data to forecast future events, but they say they make use of
more data sources—over 90 in total—which allows their system to be more
general-purpose.
There’s already a small market for predictive tools. For example, a startup called Recorded Future
makes predictions about future events harvested from forward-looking
statements online and other sources, and it includes government
intelligence agencies among its customers (see “See the Future With a Search”).
Christopher Ahlberg, the company’s CEO and cofounder, says that the new
research is “good work” that shows how predictions can be made using
hard data, but also notes that turning the prototype system into a
product would require further development.
Microsoft doesn’t have plans to commercialize Horvitz and Radinsky’s
research as yet, but the project will continue, says Horvitz, who wants
to mine more newspaper archives as well as digitized books.
Many things about the world have changed in recent decades, but human
nature and many aspects of the environment have stayed the same,
Horvitz says, so software may be able to learn patterns from even very
old data that can suggest what’s ahead. “I’m personally interested in
getting data further back in time,” he says.