Clicking with consumers
Like most people, you probably make a lot of decisions about what music, movies and books to buy based on recommendations.
Well, the Internet is trying to be your new best friend.
Dozens of companies are working on ways to create online recommendations that may one day replace -- or at least supplement -- the opinions of your pals. But as anyone who’s been flummoxed by this kind of Web-data-based reasoning -- “If you loved that John Grisham novel, you might like this bathing suit!” -- the field still has a ways to go.
As one researcher, University of Minnesota computer science professor Joseph Konstan, put it, “Amazon.com doesn’t know enough about me to recommend kitchen tools.”
But if online recommendations strike customers the right way, they might be the best thing that ever happened to more obscure music, film, literature and other art forms, not to mention the people who create them.
“It’s not really hard to figure out if [consumers] want to see the new ‘Matrix’ movie,” Konstan says. “It’s much harder to get consumers to be aware of a 1983 comedy they’d never heard of. With a recommendation system, they have the chance to find out about that.”
Retailers like Amazon.com have a record of books, music and other purchases customers make. Computer software allows that storehouse of data to be mined for comparisons and patterns -- people who bought books by William Thackeray also bought books by Charles Dickens, for example. Customers are invited to review their purchases -- liked it; hated it -- to help refine the process. That way the retailer can come up with recommendations for individual customers to consider the next time they are looking for something to read.
A leading researcher in the field, Rashmi Sinha, says a 2002 study showed “users felt that recommendations from online recommendation systems were often ‘new’ and ‘unexpected,’ while items recommended by friends mostly served as reminders of items users had already planned to pursue.”
The key is customer acceptance, and that can be deeply affected by whether the user suspects the recommendation is just a thinly disguised sales pitch or if the suggestions don’t seem to make sense.
Netflix, an online DVD rental firm, has created one of the Web’s most ambitious recommendations engines, which has helped it survive and even prosper in the post-dot-com economy. The service recently signed on its millionth subscriber.
The company charges $19.95 a month for consumers to have three rental DVDs in their homes at any one time. Titles come in the mail with postage-paid return envelopes. Netflix encourages users to rate films and will provide rental recommendations based on that information.
“We have 500 [Indian] Bollywood films, we have hundreds of PBS documentaries and we have all the mainstream hits,” says Reed Hastings, CEO of Netflix. “We provide recommendations so that consumers can find great movies they might not have heard of.”
In its attempt to woo new subscribers, Netflix is in the midst of opening a chain of distribution centers, including a new one in Chicago, that will reduce the amount of time customers have to wait to get a film they’ve selected.
But the firm’s recommendation engine, which generates 60% of all Netflix rentals, is a big part of its plan to dominate the DVD-by-mail field, recently entered by Wal-Mart and Blockbuster.
“The average subscriber is going to watch five or six movies a month,” Hastings says. “We’d like them to really love those movies.”
Netflix’s ideal scenario is a demonstration of that cliche, a win-win situation: If a consumer rents an obscure animated Japanese movie, as opposed to a popular new release that’s in short supply, Netflix wins. And if the subscriber not only likes the movie but falls in love with a whole new film genre, he or she wins too.
But churning out good recommendations is an art, not a science, those in the field have found.
“There was a lot of talk a few years ago [saying], ‘We’re going to develop a magic algorithm that knows everything about your taste, that knows everything about you. It’s going to be great,’ ” says Michael Papish, founder of MediaUnbound, the firm that supplies recommendations for the music site Pressplay. “But there is no one solution, one algorithm that gives you everything you need to know about someone.”
Cross-genre recommendations -- extrapolating people’s taste in kitchen gadgets from the books and CDs they buy -- is one of the industry’s most difficult challenges, and one that probably should be avoided, according to experts in the field. A recommendation that is too generalized or obvious is another thing to avoid, Konstan says.
“I could write a supermarket recommendation program that said, ‘Try bananas, bread, eggs and milk.’ No [supermarket] would buy that,” Konstan said, because those are obvious shopping-list items. “The hardest thing is telling you something you didn’t already know.”
Another key factor in building consumer trust, says Neil Hunt, vice president of e-commerce for Netflix, is the buyer’s belief that a site isn’t touting a particular product because it makes more money by doing so.
“Some of our products are cheaper to us than others, but we don’t let costs determine what to recommend,” Hunt says.
To be sure, different sites’ different goals should be taken into account, Papish says.
Amazon.com, for example, wants people to buy the items the site recommends, but Pressplay and Netflix want people to be so excited by the discovery of new music or movies that they stay on as subscribers.
Also, there’s no one way to build a recommendation engine, as Hunt points out.
“One of the things about movies is that the universe of titles is manageable -- there are about 15,000 DVDs that have been released ... ,” he notes. “That’s a much more manageable set than the millions of book titles or music tracks that exist.”
Even within the world of music sites, Papish says various sites go about classifying information in different ways.
“Music preferences are partly based on what the music sounds like, but [for us] it’s a lot more about the social things behind music. Who are your friends and what do they listen to? What do you think is cool? Do you listen to the radio or hate radio? But other sites will classify a track based on the fact that it has 83 beats per minute. That’s not going to help you figure out if you’re going to like it.”
His staff has gone so far as to subdivide 800 different musical genres on a gigantic map. “Our team argues about these things; it’s a fun process to watch.”
Almost as fun as getting a great movie recommendation.
