The rise of computer-driven recommendation systems designed to help consumers navigate a growing ocean of choice is prompting concerns that the hyperpersonalization of information sources will lead to harmful divisions throughout society.
New Wharton research into consumer purchasing patterns suggests the opposite. An empirical study of music recommendations indicates that receiving suggestions tailored to individual listeners actually widens exposure to new products and fosters human bonds.
In a paper titled, “Will the Global Village Fracture into Tribes: Recommender Systems and their Effects on the Consumer,“researchers examined iTunes purchases and networks that developed among people who had signed onto a service that uses a recommendation system to suggest songs the listeners might like.
The researchers found that rather than creating entrenched factions, recommender systems expand the reach of individual consumers across wider networks of people, building on common interests among listeners who expose one another to new artists. “What we found in our setting was that commonality increased among users, and people seemed to be using personalization to widen their interests and explore new things — in this case music,” says Wharton professor of operations and information management Kartik Hosanagar, co-author of the paper along with Wharton statistics professor Andreas Buja and Daniel M. Fleder, who earned a PhD at Wharton in 2009.
In reaction to recent debates over whether business and society is growing increasingly fragmented, the authors decided to take a closer look at the question using data from a recommender system — the kind of tool that sparked the debate in the first place.
The paper cites articles and books — including Republic.com by Cass Sunstein, a law professor who is now overseeing the Obama Administration’s regulatory policy — which argue that increased personalization is creating fragmentation throughout society. Sunstein suggests that personalized recommendations limit media consumption to narrow, predefined outlets, choking off exposure to the ideas and opinions of others. In The Filter Bubble: What the Internet Is Hiding from You, Internet activist Eli Pariser writes: “The filter bubble is the invisible, personal universe of information that results…. The world you see online and the world I see may be very different.”
Most of the discussion, however, has been based in theory or on anecdotal approaches, according to the Wharton researchers, who set out to conduct an empirical study that examined whether heightened personalization of the Internet is driving more, or less, commonality among individuals.
The researchers chose to focus on Internet recommender systems, which provide suggestions about products or services that might interest a consumer based on past behavior or the purchasing decisions made by others. The systems have a proven impact on consumer choice. For example, Netflix has reported that more than 60% of its rentals stem from recommendations, while 35% of Amazon’s sales originate from systems that suggest products an individual consumer might like.
“We use them every day. Netflix’s recommender for movies. Pandora’s for music. Amazon’s for books. It’s almost a part of our daily life at this point,” says Fleder. Recommender systems, adds Hosanagar, appear to be particularly important in certain product categories where there are many options to choose from, such as song titles or books.
Their paper is based on data from a music recommendation plugin for iTunes. The software suggests other songs the user may like based on titles already in that person’s music device. The technology also helps the user to sample and purchase songs that are suggested by the service, which earns commission on sales that result from its application.
The ‘Volume’ and ‘Taste’ Effects
With this data, the authors constructed a model that allowed them to examine purchasing decisions before and after receiving suggestions from the recommender system. They compared individual purchase decisions in a particular month against prior behavior. They were also able to compare the behavior of users who had received recommendations to a control group of users who, like in the clinical trial of a new medicine, had not yet received recommendations — or, in the words of the authors “were not treated.”
The study focuses on consumers who registered for the recommender service between January 2007 and July 2007. In all, 1,794 users of the service fell into the “treated” group, and 858 users were in the control group. Treated users purchased a total of 215,749 songs from 24,368 artists in the six-month period while control users purchased 106,431 songs from 14,785 artists.
The researchers examined the data further to better understand the forces driving the behavior of listeners who used the recommendation service. They found two major factors, which they called the “volume” effect and the “taste” effect.
An increase in the volume of purchases was anticipated, the authors write, but the increase of roughly 50% was larger than expected. By comparison, the number of purchases made by the control group actually declined by a small amount. “The personalization system exposes you to a lot more items you like, so you consume more than you used to before,” says Hosanagar. “As each consumer buys more, it increases the likelihood they have something in common.” For the taste effect, the study results also show that once volume is controlled for, consumers buy a more similar mix of products after receiving recommendations.
In addition to purchasing more songs, the research showed that consumers who used the service became part of networks that intensified as a result of receiving suggestions about songs. The researchers, who plotted relationships between thousands of users and millions of songs, found a 23% increase in the percent of listeners with an artist in common compared to the control group.
The researchers also plotted combinations exploring the “distance” between pairs of users, or the number of people in the network between them. They wanted to determine whether those who initially were close on the network become closer, while others who were farther removed grew farther away, indicating fragmentation. The authors found that all kinds of users — close as well as far — became closer to one another on their networks in the treated group relative to the control group. The group that received recommendations showed more user-pairs becoming closer (36%), while fewer pairs (9.2%) moved farther apart.
“The increase in similarity appears uniform: All types of users become closer to one another,” the paper states. “Users who were close became closer, and users who were initially far became closer, too.”
Water Cooler Talk
While personalization is increasingly important in business, the authors stress that it should not come at the expense of commonality. “Commonality,” says Fleder, “means being at the office water cooler and mentioning an artist. The person next to you says, ‘I really like them too! What did you think of…?'”
The ability to connect with others and to have a meaningful dialogue with them, whether at the water cooler or elsewhere, is at the heart of the debate on social fragmentation, adds Hosanagar. The Wharton study shows that the systems are not so hyper-personalized that they push down information that would create a unique information bubble for an individual. Recommendation systems capture products and information that might be on the fringe, or beyond, to extend that person’s reach into new areas. “They give you a range of experiences. It’s not so narrow that they are pushing you down a funnel,” he says. “They expose the users to different things, and the users themselves use that to discover other interests.”
Fleder notes that media tap into the essence of being human, which is a trade-off between individuality and commonality within a group. “Smart firms get the right balance between getting the right product to the right person but also keeping the human element — the commonality between you and others,” he says.
Firms need to strive for a balance between highly focused, tailored approaches to customers and those that incorporate flexibility and encourage exposure to new ideas and trends that others are exploring. “The wrong approach is matching people to products in a vacuum,” Fleder adds. “The right balance matches you to the products that are best for you, but also understands the human element — that people want to consume similar things.”
The implications of this type of shared consumption may vary by the type of product or firm, according to Hosanagar, who notes that commonality and shared experience may be more valuable in books or music than other products. Indeed, the paper points to research by Wharton marketing professor Jonah Berger that shows consumers tend to reject commonality when choosing fashion or status items.
While Sunstein and others raise alarms about a media divide that tends to limit people to their own echo chamber, the Wharton study of recommender systems may provide some new insight into the debate, according to the authors.
“We studied music. It would be interesting to extend the study for news,” says Fleder, adding that Google News, for example, provides recommender systems for individuals. “Sunstein has a valid concern — that people will only see news that reinforces their views and they won’t see information on a wide range of topics and ideas. At least when it comes to music, we didn’t find that — people actually widened their interests.”
