You’ve heard the age-old advice that it's foolish to jump off a cliff "just because everybody else is doing it." Now, researchers are turning to particle physics to prove that the most popular choice is not always the best one.
Whether Google is serving up a common search phrase as you type or other websites are supplying algorithm-based popular recommendations for a film, restaurant or hotel, the path less traveled—the product or service that is not the most popular choice—can be the most favorable one, the researchers found. In fact, they concluded that when they imposed a sort of artificial occupancy limit—so that it was impossible to recommend the same thing to a large percentage of people—a greater diversity of recommendations became available and chances that the recommendations would better suit consumers improved.
The work closely parallels the laws of physics that dictate that particles must occupy the most favorable energy states. Some particles, known as fermions, are restricted to otherwise unoccupied spaces. Other particles, called bosons, can share the same space.
"In physics, one often studies systems made from a large number of interacting components—such as atoms in a lattice or molecules in gas—where detailed modeling of individual parts is possible, but more insights can be gained by addressing [the collective group’s] average properties," says Matus Medo, a physics researcher at the University of Fribourg in Switzerland who participated in the study along with colleagues from the same university and from the University of Electronic Science and Technology of China in Chengdu, China.
"There is a close analogy between users competing for items and particles competing for states," he says. "In social systems, too, it is difficult to model and predict detailed behavior of an individual, but much more can be said about collective behavior that can emerge in a society."
The researchers discovered the unexpected success of the occupancy limits while applying them to a set of DVD film ratings, which included data from thousands of users of the Netflix online DVD rental service.
Medo says that online recommendation systems tend to be built with an inherent popularity bias, with users tending to gravitate toward items that are most popular and perpetuating this popularity cycle. The systems could be improved by adding some crowd-avoidance measures that broaden the field of recommendations, he says.
"There is this built-in tendency to concentrate the attention to fewer objects, with Google directing us to Web pages ranking high in its PageRank algorithm," Medo says. "This tendency is self-reinforcing: The more users go to highly ranking pages, the higher the rank of these pages in the future."
In the case of restaurants, a popular one can run the risk of becoming overcrowded and noisy, while a lesser-known one may provide a better overall experience for the diner.
It will be interesting to see how the findings can be applied to online engines, he says, and to further study the overall effect of popularity bias on "information ecology."
"If the bias is strong and we don't want our information and cultural horizons to narrow down too much, we might need to employ some countermeasures that automatically enhance diversity of the recommended content," Medo says.
The research has been accepted for publication in the journal EPL, formerly Europhysics Letters, and is posted on the arXiv.org website.
