They want to get inside your head

Sometimes you find a video-store clerk who knows exactly what movie you’ll like. Or a clerk in a music store who senses your taste in bands. Or a bookworm who can deliver one terrific novel after another from the shelves.

Marc Pickett wants to take luck out of that equation. And win a million dollars in the process.

Pickett, a doctoral student at the University of Maryland, Baltimore County, is trying to perfect a “recommender.” That’s a computer program designed to analyze your cinematic tastes and predict which movies you’ll like.

Online retailers, including Netflix, Amazon.com and Apple’s iTunes, put great stock in recommenders. They’re critical tools in efforts to establish long-term relationships with customers and sell more products.

The program is particularly critical for Netflix, a mail-order movie-rental giant whose livelihood depends on keeping customers happy enough to pay $5.99 or more every month.

In October, the company offered a $1-million prize to anyone who could develop a program 10% more accurate than its current recommender, known as Cinematch. A chance at that chunk of change set thousands of programmers around the country, including Pickett, to work.

“Imagine that our website was a brick-and-mortar store,” said James Bennett, Netflix’s vice president of recommendation. “When people walk through the door, they see the DVDs rearrange themselves. The movies that might interest them fly onto the shelves, and all the rest go to the back room.”

Raw computing power

Elsewhere, work on recommenders has led to prototypes of mood-sensing digital music players and a program that produces virtual, three-dimensional maps of customers’ music collections.

One hurdle to improving algorithms that dispense advice is simply having enough raw computing power to study people’s likes and dislikes. To make its predictions, Netflix’s Cinematch churns through a billion movie ratings the company has collected from its customers. The program clusters customers in groups based on how they rate movies on a five-star scale.

“The problem is to help you find soul mates,” said John Riedl, a professor at the University of Minnesota who developed an early recommender in the 1990s. “It looks for people who felt the same way you did about some movies and makes suggestions based on what other movies they liked.”

At Netflix, Bennett said, Cinematch can predict a customer’s opinion of a movie within a half a star, on average. To improve on that score, the company has released anonymous data on 100 million customer ratings so Netflix Prize contestants can test their algorithms on real people.

“We’re using four souped-up computers, but we can only feed bits of the data in at a time,” said Pickett, standing in an office cluttered with computer monitors, keyboards and robotic parts.

His computer screen displayed a map he created to visualize the movie preferences of thousands of customers. People who appear in proximity on the map, Pickett said, had similar taste in movies and thus are likely to feel the same about future movies.

The university team hopes to submit the results of its first trial by the end of the year. “Winning the $1 million would be nice,” Pickett said, “but we’re really after the bigger prize: artificial intelligence.”

As of last week, the front-runner among the 13,487 competing teams had improved on the accuracy of the Netflix algorithm by 6.11%.

‘Musical primary colors’

Pandora.com, an online radio station based in Oakland, uses a different approach to pick music. Pandora staff members analyze about 10,000 songs a month for 400 musical qualities such as “subtle use of vocal harmony,” “pop rock qualities” and “major key tonality.”

“We’ve broken the songs down into musical primary colors,” said Timothy Westergren, the company’s founder. “All of those things form a kind of musical DNA.”

When listeners type the name of a song or artist into a form on Pandora’s website, it searches through a database of 500,000 titles to create a playlist of songs that share its qualities. Listeners can rate or skip songs as they play through the website. This provides feedback Pandora uses to decide what to play, Westergren said.

One drawback of these “content-based” systems is they don’t always capture musical elements important to some customers, said Paul Lamere, a computer engineer at Sun Microsystems. “If the main reason you like music is how much cowbell is in a song, but the system ignores cowbell, you’ll never get a good recommendation,” he said.

Too busy to tell a recommender what you’d like to hear? Sandor Dornbush, another graduate engineering student at the University of Maryland, Baltimore County, is working on a mood-sensing MP3 player to free you from that task.

The portable player, for now dubbed the XPod, will monitor physiological signs such as heartbeat and skin temperature to determine what kind of music to play, Dornbush said.

“It could find, for example, that when you run you like upbeat music,” he said.

But like all recommenders, human or digital, Dornbush’s prototype needs to spend time with a person to get to know him.

“If you don’t give them any feedback,” Dornbush said, “they have a tough time figuring out what you like.”