Netflix relying on big data for survival
Personalization of content for users vital to survival
Many companies toss around tech buzzwords such as algorithms, machine learning or big data, but few embed them deep enough into their DNA that their business depends on them.
Netflix Inc. is one such company. It needs big data and algorithms to survive since it depends on getting more people to watch ever more programming, which means it must serve the most appropriate shows or movies for each of its more than 93 million subscribers. Otherwise, the company said, people will quickly move on to another activity to fill their time, threatening what has become a US$62-billion business.
“We (use) all of the information we have: what people watch, when they watch, how much do they watch, what time of day, on what device, what they watch (before or after) and on what profile,” said Todd Yellin, Netflix Inc.’s vicepresident of product innovation, during a media briefing.
“Some of that data is junk, some of it is gold and we figure out how to leverage that data to put the right content in front of the right people at the right time.”
Yellin’s team specializes in making the discovery process easier for users by adding new features that reduce friction for people to get into content faster. “We are addicted to the methodology of A/B testing,” he said. “We run over 200 tests a year. And until something shows green, that the people are getting more value for their money by streaming more hours on Netflix or sticking around Netflix and retaining better, we don’t launch a feature.”
As an example, Yellin points to a new content rating system for users that will soon be rolled out globally to replace Netflix’s longstanding five-star methodology, which the company has discovered has several flaws.
For one thing, there were too many steps involved in the rating system and subscribers tended to rate something they liked more often than things they didn’t, skewing the results to the positive side. The company said it’s received more than 10 billion five-star ratings.
“We made ratings less important, because the implicit signal of your behaviour is more important,” Yellin said. “We try to measure how important it is when you click ‘play’ on a title and watch for 20 minutes versus if you watched and binged for six hours.”
As a result, Netflix will give subscribers a thumbs-up, thumbsdown rating system, which it believes is easier, quicker and involves less thought. The company began testing it last year and ratings increased more than 200 per cent.
“The most important work I think we do is around personalization,” said Reed Hastings, the company’s chief executive, during a Q&A session. “This idea that the more you watch, the more Netflix learns your tastes. Personalization is really the thing that the Internet can do that linear (distribution) can’t, and that’s a real breakthrough.”
Netflix’s algorithms look at more than just an individual’s history when it comes to serving recommendations. The company said it also tries to contextualize suggestions based on regional preferences as well as what it calls global “taste clusters.”
For instance, someone who watches a lot of action flicks will get more suggestions in the same genre, but Yellin said the taste clusters help the algorithms to also recommend unlikely titles in completely different genres.
“We have over 1,300 taste communities,” he said. “We’re finding these clusters of people and then we’re figuring out who’s like you, who enjoys the same kinds of things, and then we’re mixing and matching those.”
Netflix is using the large amount of data it collects to also introduce another new, machine-learning feature called “percent match.” Like popular online dating websites or mobile apps, the feature will use personalization algorithms to match people with a TV show or movie.
“We’re trying to create our own love story between people and content,” Yellin said.
In addition to receiving recommendations for titles with a higher match percentage, subscribers will be able to see how strong the match is on a specific title, unless the match is below 55 per cent.
“Are we perfect at this? Far from it. If we were perfect at this, we’d show you one title whenever you come to Netflix and we’d be sure that’s the one you want,” he said.
“So we try to be transparent with you. The only things we’re not transparent about are sometimes the secret sauce in our algorithms and machine learning and what we do there.
“We invest a lot in them and that’s proprietary.”
Some of that data is junk, some of it is gold and we figure out how to leverage that data to put the right content in front of the right people at the right time.