Before Netflix got into the business of producing its own programming, it spent a lot of time emphasizing its recommendation software, the algorithms that would learn your taste and suggest the perfect title from the company’s catalog. Over the last few years that feature faded into the background, overshadowed by original content, Emmy nominations, and its rapid global expansion.
Today, however, Netflix recommendations step back into the spotlight. Up until now the recommendations you got were based on where you lived. People in the US saw suggestions for action or comedy flicks based on what other people in North America enjoyed. Going forward, an anime fan in Sweden will see recommendations based on the viewing habits of anime fans from around the world, and the same principle will apply to every category of suggested films. This sounds like a relatively minor change, but it’s actually a precarious shift for a core technology that the company has been putting off for some time, one that was a year in the making and involved a team of nearly 70 engineers.
The regional approach Netflix had been using was a stop-gap solution to prevent local catalog differences from throwing off the recommendation algorithm. "We were very worried that running the algorithms we knew worked well when we pulled data from a single country and a single catalog, if we tried across places where the catalog differed, the recommendations would be pretty bad," says Carlos Gomez-Uribe, vice president of product innovation at Netflix, and the leader of the recommendation redesign. A pair of Netflix engineers explained the problem at length in a blog post published today:
The dystopian Sci-Fi movie "Equilibrium" might be available on Netflix in the US but not in France. And "The Matrix" might be available in France but not in the US. Our recommendation models rely heavily on learning patterns from play data, particularly involving co-occurrence or sequences of plays between videos. In particular, many algorithms assume that when something was not played it is a (weak) signal that someone may not like a video, because they chose not to play it. However, in this particular scenario we will never observe any members who played both "Equilibrium" and "The Matrix". A basic recommendation model would then learn that these two movies do not appeal to the same kinds of people just because the audiences were constrained to be different. However, if these two movies were available to the same set of members, we would likely observe a similarity between the videos and between the members who watch them. From this example, it is clear that uneven video availability potentially interferes with the quality of our recommendations.
The regional model, however, created a new set of issues. Each time Netflix launched in a new territory, it was trying to make suggestions without knowing anything about its customers. To work well, the software needed a massive data set to crunch. Without that, recommendations would be heavily influenced by the handful of users who interacted with the service. If a few oddballs among the first dozen people who signed up in Madagascar happened to watch The Notebook one night and Old Boy the next, well, that nation’s recommendation system would think the movies are a good match. To get around this, Netflix's recommendation team basically had to do the work typically reserved for its machines. "When we launched in new markets — Sweden, Germany, France, any of our previous markets — we had to spend a lot of time hand-tuning the recommendations," says Gomez-Uribe.
Once it became clear internally that Netflix was going to vastly expand its territory, the recommendation team decided to bite the bullet and move off regional recommendations. The team spent a year creating a global recommendation system that leverages the taste graph of the entire subscriber base while simultaneously taking into account differences in regional catalogs.
It was time to bite the bullet
The new system has actually been live since the global expansion rolled out at CES, although it wasn’t publicly announced then. When the first dozen customers signed up in Macao, for example, they were fed recommendations that weigh their initial selections against what customers around the world have already established. If they checked out documentaries on food, they were then directed to a feed of foodie films that reflects what culinary cineastes enjoy from Miami to Moscow. For regions that already have an established taste graph, Netflix will now layer that local flavor on top of the global recommendations, and this will happen for every country once they generate enough data for the algorithms to produce accurate results. "Even when we launch on a tiny island, that first person who signs up, the moment they start browsing through movies, we can begin to offer them more personalized recommendations," says Gomez-Uribe.
The next big challenge for Netflix is language
The next big challenge for Netflix is language. Today Netflix is available in 21 languages, but doesn’t specifically gather information from users on their native tongue. "Currently which languages a member understands and to what degree is not defined explicitly, so we need to infer it from ancillary data and viewing patterns," Netflix engineers wrote in today’s blog post. When ranking recommendations, should Netflix show the best match, or the best match in the language that is most comfortable for that user? Today’s announcement doesn’t include any new recommendations based specifically around language, but Netflix says that it does have live tests running in the field as it works on optimizing its algorithms for this problem.
I asked Gomez-Uribe if the new recommendation algorithm would make any attempts to tune its suggestions based on cultural differences. How would a comedy about Hitler play in Spain versus Germany versus Israel, for example, and how might that affect recommendations. He demurred, saying, "We are very aware that different cultures perceive nudity and violence in different ways. Part of our challenge is we don’t want to be in the place of guessing how a culture will interpret something." Netflix will probably offend a few customers and generate a few headlines when the algorithm runs awry. But those are problems borne of success. "Countries have different attitudes to what content is appropriate," said CEO Reed Hastings when announcing the expansion at CES this year. "The global potential is both a joy and a challenge to fulfill."