As part of its efforts to better compete with Google, Microsoft is plumbing the connections between searchers and their contacts to produce better results.
Microsoft researchers are exploring whether using data from several members of a social group–a technique that the company calls “groupization”–can improve search results. Their initial findings, based on experiments involving around 100 participating Microsoft employees, suggest that tapping into different types of groups could produce significantly better search results.
The team has developed an algorithm that, on average, pinpoints at least one search result for all members of a group that they judge to be better than the results returned using conventional algorithms. The results will be presented at the Web Search and Data Mining Conference in Barcelona in mid-February.
The Microsoft team believes that the approach could help the company overcome an industry-wide plateau in the quality of search results. “Today, search engines are really challenged and are sort of at the cusp of having to know individuals better,” says Jaime Teevan, a computer scientist at Microsoft Research and lead author of the paper. “This [research] has the opportunity to enrich that.”
The new research is part of Microsoft’s efforts to erode Google’s massive lead in search. Google currently attracts 63 percent of all searches, according to a 2008 survey by consumer-analysis firm Nielsen, far outpacing both Yahoo’s 17 percent share and Microsoft’s 10 percent share. Last year, Microsoft attempted to increase its share by acquiring Yahoo, but its initial advances were rejected. Yahoo later wanted to return to the bargaining table, but for the time being, Microsoft is focusing on increasing its audience by enhancing its own search offering.
With an eye on refining search results, Teevan and her colleagues–Meredith Morris and Steve Bush–looked at the way that people with similar interests or attributes search for information. The researchers grouped people using explicit factors, such as their age, gender, participation in certain mailing lists, and job function. In some cases, implicit groups–such as people who appeared to be conducting the same task or appeared to have the same interest–were inferred. The researchers acknowledged that gathering such data in the real world could be tricky. But it could perhaps be collected through registration, by caching previous searches or by tapping into social-networking software.
The Microsoft team found that groups defined by demographics such as age and location have little in common for most searches. However, groups of people with similar interests tend to rank similar search terms highly. The researchers also found that, while people believe that they phrase their queries in similar ways, the idiosyncrasies of search terms vary tremendously.