Ten years ago newly hired mathematical physicist Jennifer Chayes told her boss, Microsoft Corp founder Bill Gates, about new methods, derived from the phase-transition theory of spin glasses, to solve constraint-satisfaction problems in social and other networks. She warned him, however, that they “would take 100 years to pay off.”
Chayes, who cofounded Microsoft’s theory group with her husband and fellow physicist, Christian Borgs, recently contacted Gates to say, “I can’t believe it, Bill. It has only taken 10 years to pay off.”
Now, as director of Microsoft’s new research laboratory in Cambridge, Massachusetts, Chayes will assemble and lead groups of social, computer, and physical scientists to model and design online social networks. Microsoft Research New England is the company’s sixth research lab and the first with the mission of bringing together social and computer scientists to work on algorithms for social computing applications. More than just message boards, online social network applications such as Facebook, MySpace, and LinkedIn have become popular venues for advertising and search companies; industry analysts speculate that the new Microsoft lab and the company’s bid for internet rival Yahoo Inc point to the urgency that the software giant is placing on competing online. Microsoft Research New England is expected to open this summer, just less than a year after it was first proposed, says mathematician Henry Cohn, a founding member of the new lab. “One of the advantages of industry is that when there is a compelling case for something, it can get done quickly.”
Use of the Web has surged with the popularity of so-called Web 2.0 applications, which allow users to generate content and form communities. Several physicists, including Chayes, saw opportunities in the late 1990s to apply statistical mechanics principles to analyze complex networks like the Web. Peter Norvig, director of research at Google Inc, says the online search and advertising company employs “well over 100 people with one or more degrees in physics” to work on mathematical problems. “The reason I think that physicists do so well [in network theory] is that we are used to dealing with very large systems with lots of similar and interacting entities,” says Chayes. “In the case of the World Wide Web, for example, there are on the order of 100 billion static webpages and even more dynamically generated webpages.”
Highly connected hubs
Researchers studying self-organizing social networks look at how links are formed between individuals, whether some individuals or nodes are better connected than others, and the collective action or behavior of the entire network. In the past social scientists relied on surveys and questionnaires, but on the Web “social behavior is self-documenting—it leaves traces behind,” says Microsoft research sociologist Marc Smith, who studies and designs improvements for social online applications.
Duncan Watts, director of the human social dynamics group at Yahoo Research, is using Friend Sense, an application he wrote for Facebook that queries users’ political attitudes and how well they know their friends. Watts, who holds a bachelor’s degree in physics and is also a Columbia University sociology professor, may be best known for showing that so-called small-world networks can be characterized by a small path length between nodes (related to the six degrees of separation) and a large degree of clustering among nodes (see Physics Today, Physics Today 0031-9228 51 9 1998 17 https://doi.org/10.1063/1.882433 September 1998, page 17 ). For example, Bernardo Huberman, a physicist and director of the social computing group at Hewlett-Packard Co labs, recently scoured 362 million messages, minus the identifying information, of 4.2 million Facebook users and found that college students cluster by school affiliation.
“Social networks do offer one of the best-described systems that you can monitor in the most precise way,” says Albert-László Barabási, director of the Center for Complex Network Research at Northeastern University. “We are using them to find fundamental organizing principles that can be tested in other systems as well.” In 1999, Barabási used Web-crawling data to show that many complex networks were scale-free—a few nodes are highly connected hubs. Models that involve universality laws for complex networks are being applied to studies of how viruses spread on the internet and in human and biological populations, among other things, says physicist Mark Newman of the University of Michigan Center for the Study of Complex Systems. Online advertisers and developers can also take advantage of such network models to tailor their services for users based on interaction patterns or even devise incentives and mechanisms to influence people’s behavior, says CSCS physicist Lada Adamic, who with Huberman in 1999 also showed that Web growth is scale-free.
Spin-glass payoff
Some condensed-matter physicists are drawn to social network modeling because it is similar to a many-body problem, says Huberman. Like spin-glass materials that have disordered and unpaired magnetic spins, individuals have conflicting interactions with their neighbors, and their uniqueness leads to disorder, says Université de Paris-Sud physicist Marc Mézard. It’s a patent from Mézard’s spin-glass theory work that is now paying off for Microsoft: He, Chayes, and collaborators are using that patent to solve optimization problems such as sending messages from one node to others, bypassing intermediates.
The researchers interviewed for this story say precautions are taken to protect the privacy of the personal information they use. “There’s a very sticky privacy issue here,” says Yahoo Research head Prabhakar Raghavan. “Data that we get from [our academic collaborators] goes through all manner of scrubbing and approval procedures within the company from the legal department. The principle we follow is, we use the data as we need it, then it gets destroyed.”
Complex network models do not capture the nuances of human behavior, says Huberman. “People, unlike atoms and spins, are intentional beings, which limits the validity of many-body physics approaches.” Sometimes individuals do alter their behavior, adds Barabási, “but at the end of the day, collective behavior doesn’t change.”
A network of coauthors among physicists who had published papers on networks as of 2003 shows the formation of discrete communities, denoted by colored nodes; the size of nodes corresponds to the size of the community. The thickness of the links connecting the nodes is proportional to the number of pairs of collaborators between communities.
A network of coauthors among physicists who had published papers on networks as of 2003 shows the formation of discrete communities, denoted by colored nodes; the size of nodes corresponds to the size of the community. The thickness of the links connecting the nodes is proportional to the number of pairs of collaborators between communities.
