"Search in many ways is brute force," Dumais said. "If the two of us type in a query, we get the same thing back, and that is just brain dead. There is no way an intelligent human being would tell us the same thing about the same topic."
Personalisation was one of the big buzzwords of the early years of the dot-com era, but many of the efforts to deliver individualised content failed. Software developers, however, are increasingly becoming more adept at using Bayesian models and other probabilistic techniques to insert intelligence into software.
Although the underlying calculation in these models is complex, the overriding concept is fairly simple. Software keeps tabs on an individual's Web surfing habits, interests, acquaintances, work and travel history, work projects, and other data. It also constructs a model that tries to anticipate what a person finds important and what will be irrelevant.
"I have the same meeting every week with the same people. Maybe that isn't so important," Dumais said. "I have a meeting with Bill G. [Gates] He's pretty high on the org chart. Maybe that one is important."
Microsoft's experiments differ from commercial search engines in that the universe of data searched consists of data found on an individual's hard drive. Although a smaller universe, it's a well-travelled one. Studies cited by the company suggest that up to 81 percent of Web pages accessed are repeat visits. Hence, the links someone wants to see are likely on his or her hard drive.
There is also no theoretical reason the scope of these type of searches couldn't be extended, which would allow Longhorn or other search-enhanced applications to compete with commercial search engines. Dumais pointed out that search queries could take into account the geographic location of the PC used in the search.
While Google and other search engines also rely on probability, the results typically reflect the preferences of the Web surfing population at large. Microsoft is essentially trying to bring it more to the individual level.
