A New Way to Know What Users Really Want Online
Research by Shawndra Hill and Olivier Toubia, Marketing
Two Columbia faculty have developed a new model for understanding how people respond to search results, one that could help search engines and advertisers offer users more finely targeted content. Olivier Toubia, the Glaubinger Professor of Business and chair of the Marketing Division, and Shawndra Hill, principal scientist at Facebook and senior lecturer in the Marketing Division, along with Jia Liu, PhD '17, of Hong Kong University of Science and Technology Business School, analyzed the search tendencies of Bing users during the 2016 Super Bowl in order to design what they call a flexible content-based search model.
In their paper, “Content-Based Model of Web Search Behavior: An Application to TV Show Search,” the researchers detail how people search for information throughout the course of a TV event. The study explores query-usage dynamics, looking at the distribution of certain search words before, during, and after the Super Bowl.
Some queries were used much more frequently, “suggesting different overall levels of consumer interest across Super Bowl topics,” they write. Some queries, such as “Super Bowl” and “Carolina Panthers” were searched for consistently over time; and others, such as “kickoff time” and “prediction” exhibited usage dynamics, meaning they cropped up more or less frequently before, during, and after the event.
The study found that even when queries remained constant and a user entered a generic search term, like “Super Bowl,” the links they clicked on changed depending on whether they were searching before, during, or after the event.
“Not only were people searching differently for these topics, but they were also clicking differently,” says Hill. “They were searching for the term ‘Super Bowl,’ but before the TV show aired, they were clicking on things like what time the event was on, versus after, when they clicked on things like who the MVP was.”
The researchers used the click dynamics and topic modeling to better predict how and when users would click on certain search results.
Toubia, Hill, and Liu are aiming to fill the gap in this important information by developing a model that identifies and quantifies users’ content preferences across search contexts and can predict click-through rates on the search engine results page. “What we wanted to understand is those dynamics and also, whether we could jointly model the search behavior and the click action in a way that would help us to be better at predicting what people want to see and what they, as a result, want to click on,” says Hill.
—Sara Cravatts