Can an Algorithm Know Which Movie You Want to Watch?

Picking a movie to watch on a streaming service isn’t complicated; you point and click on a film you might find entertaining. But according to Olivier Toubia, the Glaubinger Professor of Business, the psychology of why you chose that movie could be more than a simple matter of taste.

“Is it because there is a deeper reason?,” Toubia says. “If we can put our finger on that, then streaming services might be able to be more systematic and broad in their recommendations.”

And seeing that services such as Netflix and Hulu are acutely interested in what people are watching, there’s value in discovering how content affects their customers’ viewing choices.

Two Competing Recommendation Models

In a paper published this year in Journal of Marketing Research titled “Extracting Features of Entertainment Products: A Guided LDA Approach Informed by the Psychology of Media Consumption,” Toubia and co-authors Garud Iyengar, professor of Industrial Engineering and Operations at Columbia University, Renée Bunnell of Emoto, and Alain Lemaire, a PhD candidate at the Business School, illustrate the potential of new recommendation models that could be used to predict movie-watching behavior.

According to Toubia, previous scholarship on recommendations usually involves two different approaches. The first, called the “collaborative approach,” matches films with viewers based on the viewing history of other similar viewers. Toubia says it is the “standard way of doing it,” but his team found the method has limitations.

“It’s not as powerful when you have a new movie that no one has watched yet,” Toubia says. “Also if there is a new a Netflix user, we have no idea what they like and they have no history of movie-watching to match with other users.”

The other approach, “content based,” tries to quantify how much each consumer likes different movies or television shows based on their own past behavior. “For example, I might have a movie that no one has seen yet, but I know the film features a specific type of content,” Toubia says. “Then I can try to guess how much you might like this movie based on how much you like movies with similar themes.”

As Toubia notes in his study, the research team developed a better content-based method to gather input that is “objectively defined, predictive of consumers’ decisions, and not excessively complex.” “The traditional way involves using genres—comedy, drama, etc.,” Toubia says. “But genres are not enough to exactly describe a movie; there could be two comedies that are very, very different.”

Applying Principles of Psychology to Content-Based Algorithms

Toubia and his co-authors rooted their study in two branches of psychology—media psychology, or the psychology behind the consumption of entertainment, and positive psychology, which tries to help people lead more meaningful and fulfilling lives.

Developed by psychologist Martin Seligman, positive psychology contains a taxonomy of 24 character strengths such as social intelligence, love, kindness, and perseverance that exist for every individual. “Maybe you are someone who’s very strong on kindness or humor,” Toubia says. “If you nurture those character traits, you would tend to lead a happier, more meaningful life.”

Toubia and his co-authors sought to improve the quality of the recommendations by looking beyond genres to focus instead on the number of themes that correspond to these character traits present in a movie.

In order to identify themes, Toubia and his co-authors used a topic generation model—Latent Dirichlet Allocation (LDA)—which finds groups of words that tend to appear together, such as “police,” “cars,” “chase,” and “guns.” The research team then “seeded” the LDA model with 96 different words that link to various character traits.

“At the same time, we know that there are movies about many topics,” Toubia says. “So, the method we developed is flexible in that it finds these themes that maybe don’t have anything to do with psychology, but are still relevant to why people choose certain movies.”

Check out Ivy Exec’s recommended reading list for CEOs and rising business leaders.

Creating an Intuitive User Experience

Using this expansive LDA model, Toubia and his co-authors discovered that people don’t always select movies just because they embody the viewer’s values or morals; people are also drawn to films that demonstrate the opposite of a viewer’s character traits.

In the paper, Toubia discusses the 2009 comedy, The Hangover, about four men at a bachelor party in Las Vegas that goes terribly awry. “The movie is about careless people doing crazy stuff,” Toubia explains. “There’s a theme of self-control, but actually the movie shows the opposite of that.”

The research indicates that making meaningful recommendations is not as simple as matching profiles with content. It is about trying to understand, for each consumer, what type of content each consumer wants to watch. “It is a way to describe content along certain dimensions and then quantify how much each consumer likes content that features each of these dimensions,” Toubia says. “It could potentially lead to recommendations that are more meaningful and not just enjoyable.”

Read the original piece on Columbia Business School’s Ideas and Insights blog. 

Ivy Exec is proud to announce its partnership with Columbia Business School, to bring an insightful collection of thought leadership pieces for the modern-thinking strategist in finance, leadership and more to its platform.