Yahoo's COKE Project Adds Life to Front Page

Yahoo could add millions of dollars to its bottom line using a new artificial intelligence system to figure out what news stories its readers like, a company researcher said Thursday.

In the past few months, Yahoo has been testing a system that helps it identify the most elusive aspect of a story: whether it is hot or not.

Yahoo Senior Researcher Deepak Agarwal is part of a team of about 10 researchers that has spent the past few months testing whether their artificial intelligence algorithms can help boost site traffic within the Yahoo Today module -- the part of the Yahoo.com page that features the top news stories of the day.

Internally, they call the system the Content Optimization Knowledge Engine (COKE), and the geeks who work on the system jokingly call themselves Cokeheads.

Yahoo's editors already have pretty good instincts about whether a story is hot, and convincing the people in charge of the front page to let machines make calls about story placement was initially a challenge, said Agarwal, speaking at the New Paradigms in Using Computers conference at IBM's Almaden research facility. "Editors have been doing this for years," he said. "Convincing them that this would work was not easy."

But the researchers have shown how the system can help humans better understand what's going on with the Web site. They've created a dashboard that can instantly show editors how certain stories are playing among different groups -- young adults versus young boomers, for example.

To solve the hotness problem developers had to go beyond the usual approach to selecting stories, he explained, instead of basing suggestions solely on factors such as age, gender and location of the reader, and keywords in the article. "We tried to personalize, sorted by age and gender, but it did not work," he said.

That's because this approach didn't take into account things such as the fact that certain types of stories are more likely to be clicked at certain times of the day.

So researchers developed a model that takes a variety of factors into account and measures how popular stories are doing almost instantaneously. "We had to build a dynamic model," Agarwal said.

Collecting and crunching all of this data quickly enough for it to be relevant in a world where stories grow cold after six hours is the hard part, he said. Building a system that could be used even when traffic demands increased was also a top priority, he added.

It's still not clear when, or if, COKE will leave the research phase, but the early results show promise. COKE has boosted click rates by 25 to 30 percent, Agarwal said. That could translate to several millions of dollars in revenue per year.