How Netflix knows you'll love 'Step Brothers' but loathe 'Avatar'

The company's vice president of personalization technology gives us a behind-the-scenes look at the mechanics of the ever-evolving Netflix brain.

Ever wonder how Netflix accurately predicted that you'd love The Secret Life of Bees, even though you were skeptical?

The answer: A lot of manpower -- and even more algorithms. About one-third of the company's 500 or so engineers work specifically on personalization technology. That is, knowing your kids will love the movie Up based on your high ratings of Ratatouille and Toy Story. And in addition to the salaried personalization engineers, thousands of others contribute ideas through the Netflix Prize.

I spoke last week with John Ciancutti, vice president of personalization technology, who gave a behind-the-scenes look at the mechanics of the ever-evolving Netflix brain.

What's involved in the process of linking Netflix users with movies they'll enjoy?

In each format [DVD and streaming], we have thousands and thousands of choices that we offer each of our customers. Instead having one website where you and I both see the same movies to choose from, we have 14 million different websites, a different website for every customer. The power of personalization helps us choose which of those movies they're going to be most excited to see. The science we apply to this problem takes a variety of forms. You may be familiar with the Netflix Prize. That was an opportunity for the academic community to improve the mathematical models behind our movie prediction engine. We also use a variety of other techniques to try to understand what movies are like other movies, what are the ways customers use our service and our system and how can we use that perspective to improve the quality of the movies we're showing them.

How has the system evolved over time?

This system has evolved constantly over the last 12 years. We have new ideas, new algorithms, new features, new user experiences that we want to bring to our customers. We'll try those [new experiences first as] tests. We'll give those new experiences to some customers, but not to all, and we'll measure the impact that new idea has on those customers. If they're finding more titles they're excited to see, if they're loving the movies they see and if they're rating those movies more highly, then we'll know that idea was a successful one. That will teach us more about what our customers want and we'll roll that idea out to our customers. There are literally dozens of these tests going all the time. It's a relentless and constant innovation exercise.

What, specifically, have you learned about how to make personalization better?

Most of our customers rate movies on our website and on television-connected devices. We've worked very hard to try to understand how best to use those explicit taste ratings in order to improve the experience we offer. We have three billion movie ratings. Those ratings don't just help the customers that give us those ratings. They help us understand movie ratings generally and in that way can help all 14 million members. We've learned over the last few years how we can use implicit data that our customers give us, as well. By implicit data, I mean how are the ways they use our service? What movies do they choose to stream? What movies do they choose to ship on DVD? What movies do they add on their DVD or streaming queues? What does that data tell us about their taste? How can we leverage that data to give them a better experience than we did before?

What's the most important data you get from users?

Our customers really engage with our service. They use us a lot. They're on our website a lot. They use our applications with television-connected devices very frequently. The best data for us is that combination. We can combine the movies you say you most loved in the past with the movies you're choosing to stream right now with the categories and genres you tell us you want to see. If we have a customer that rates movies for us, but they don't rate genres or categories and maybe they haven't rented any movies from us yet, we'll do our best to leverage their ratings data. Conversely, if they haven't rated movies, but they've streamed a lot of content to their Xbox or their Nintendo Wii, we'll use that data as best as we can, too.

What are the primary challenges you face as you try to improve the system?

It's a challenge of innovating. We've only scratched the surface of what we think personalization is capable of. We think that decades from now, people will be working very hard on personalization and how to make media choosing better. Our challenge is to be one of those leaders decades from now by continuing to be a leader today. The pace of innovation we think we need to achieve in order to provide a better experience for our members is very high. That's why we take so many different approaches to how we do innovation. We have algorithm teams that are focused on the latest mathematical techniques. We have teams of movie and television show experts that create different categories of movies. We have design and user engineering teams that are focused on creating experiences around those television show. Generally, our challenge is to innovate relentlessly and effectively and out-pace all the other ways that people can get movies and television shows.

Do you still face the Napoleon Dynamite problem mentioned in a 2008 New York Times Magazine article?

There are a lot of challenges that we're working on now. What was called in that article the Napoleon Dynamite challenge is certainly a perennial challenge in the space of movie and television show recommendations. There are titles out there like Napoleon Dynamite that just sort of defy conventional definitions. It's very hard to predict if customers will like them. For those particularly challenging movies, we use a combination of factors to figure out who the people are that we're going to recommend this title to. On one hand, we'll use that prediction I talked about. On the other hand, we'll use those category preferences. The thing about Napoleon Dynamite is it's a quirky film and it's a witty film. We might think you like that movie, but we might also want to make sure that we believe you like both goofy and witty movies and television shows before we recommend it to you. Let's say we know you recently streamed Punch-Drunk Love or Sunshine Cleaning. We think of both of those as quirky movies. Let's say also you've rated The Big Lebowski really highly. And let's say you also streamed Dr. Horrible's Sing-Along Blog. Those are both witty pieces of film as far as we're concerned, so that would give us some indication that not only do we think you'd enjoy the film, but we think you really like quirky and witty titles. That will give us enough confidence to suggest that you'd really enjoy Napoleon Dynamite.

How do you deal with shared accounts?

With each account we're really modeling the tastes and interests of the people who use that account. There are lots of different techniques we use in personalization to try to show movies to that particular account. We won't just show critically-acclaimed, visually-striking dramas, but we'll also show thought-provoking, political documentaries. What we try to do in those cases is not only find the content that's uniquely exciting and interesting to that particular household, but also show a range of the types of content that have been watched and loved in that household to maximize the chances that everyone who might use the service is finding content that they're excited to see.

Image, top: Netflix mailer / Courtesy of Netflix

Image, bottom: John Ciancutti / Courtesy of Netflix

This post was originally published on Smartplanet.com