The outcome of the U.S. presidential election could be predicted to a high degree of accuracy using an algorithm to analyze the social networks of "celebrity" Twitter users, according to a group of Israeli researchers.

The study was carried out by professors Zvi Lotker and Chen Avin of the Communication Systems Engineering Department of Ben-Gurion University, together with Yvonne-Anne Pignolet, a colleague from Switzerland's ABB Corporate Research Center, and Itzik Turkel, a doctoral student in their department.

These are not necessarily celebrities in the usual sense of the word, like Justin Bieber or Kim Kardashian, Lotker told Haaretz, but rather experts in various fields and other key figures. Their premise was that key tweeters wield great influence on the opinions of their followers on the social network, at least in the United States. "It's the media that truly influences the elections. And who creates and disseminates the media? The celebrity tweeters," Lotker says. The study grew out of the 2008 U.S. election, in which technology in general and social networks in particular, such as Facebook and Twitter, played an unprecedented role. Lotker said it all started when he read an article suggesting that President Barack Obama's success on social networks was based to a great extent on a huge database. "We wanted to know how the elections would look in another 10 or 20 years," Lotker said.

The researchers mapped the political opinions of a core group of tweeters and attempted to extrapolate from them who the larger group of major tweeters would vote for. Lotker said the first step was to make a list of 164 celebrities who had made their political opinions known over Twitter.

The researchers then divided that group into two smaller groups: one of 114, used to train the algorithm they had developed using language processing and machine-learning techniques, with the remaining 50 set aside to test the accuracy of the algorithm.

They then applied the algorithm to the tweets and links of the other "elite" tweeters to identify political opinions - Republican, Democratic or unknown. The researchers say they were able to peg the political preferences of the group with between 90 percent and 97 percent accuracy. Lotker says the most interesting aspect of the study were the undecided voters among the tweeters, or those who decided not to reveal their opinion. They could be the fulcrum, especially in such a close race as the current one. According to Lotker, if the algorithm could be used to identify the opinions of these individuals it could be a tool of paramount importance to parties and politicians trying to focus their messages on the people they can influence.

The graph the researchers show on their website shows, for example, how those "floating votes" appear as a barometer of public opinion. After the first presidential debate, of which Republican candidate Mitt Romney was declared the winner, the number of people the algorithm identified as Republican by their tweets spiked, the "unknowns" declined and Democrats remained more or less unchanged.

The latest study was based on an earlier study by the researchers that identified some 6,000 major users of Twitter with at least 10,000 followers.