Netflix, the company that once used to ship you DVDs, is making a smart transition to being the on-demand media streaming provider, and is also taking on cable & network television with their own productions. Over 27 million US streaming customers and the brilliant House of Cards series prove they are definitely on the right path. What makes them different from your average network or television streaming service? Recommendations, that drive an estimated 75% of viewer activity.
Search is what happens when your recommendation system fails
“Almost everything we do is a recommendation. I was at eBay last week, and they told me that 90 percent of what people buy there comes from search. We’re the opposite. Recommendation is huge, and our search feature is what people do when we’re not able to show them what to watch,” engineering director Amatriain explains to Wired.
This idea, that when a user is forced to default to search, your recommendation system has failed him, is a brilliant thought to keep in mind. In a time where customers can choose between hundreds of TV shows, there’s probably nothing more important than serving them the ones right for them on a silver platter. Because an overload of choice is never a good thing – we watch series to relax, not to have to think too hard about which ones we’d enjoy to watch, and because what we think we’ll like isn’t necessary what we will actually like. Know thyself and all that. As Carlos Gomez-Uribe, VP of product innovation and personalization algorithms at Netflix describes it to Wired: “A lot of people tell us they often watch foreign movies or documentaries. But in practice, that doesn’t happen very much.”
What are Netflix’ recommendations based on?
So if it isn’t on our claims of what we like, what are Netflix recommendations based on then? Rather than the old-fashioned user ratings, nowadays, the company is using a complex ecosystem of algorithms which they let loose on the ‘big big’ data generated by their customers, and that analyse the hundreds of TV shows themselves. In particular though, these algorithms are still looking for similarity:
- Similarity in shows, or Content-based Filtering – Netflix analyses the shows’ metadata to find similarities such as having actors in common, having the same director or genre, the year they were produced, .. . Determining show similarity is not solely a job for machines and their algorithms, but also for humans who hand-tag related shows.
- Similarity in user behaviour, or Collaborative Filtering – Which shows you search for, and which ones you actually watch, at what time of day, on which device, … is all taken into account. Even your browsing and scrolling behaviour is tracked, as are – probably – pauses and fast forwards. In general, the Netflix algorithms assume that similar viewing patterns represent similar user tastes, which is why the multiple profiles per account recently announced are of such importance.
But what about context? As Netflix has one of the most famous – and successful – large-scale hybrid recommender system currently existing, to what degree are their recommendations context-aware already?
We have been working for some time on introducing context into recommendations. We have data that suggests there is different viewing behavior depending on the day of the week, the time of day, the device, and sometimes even the location. But implementing contextual recommendations has practical challenges that we are currently working on. We hope to be using it in the near future.
Xavier Amatriain, engineering director at Netflix, quoted in Wired (August 2013)
I can’t wait for more context-aware recommendations (or for Netflix coming to Belgium, as a matter of fact). Even merely looking at a persons upcoming events and schedule could return ‘magical’ results. Imagine your TV knowing you have a meet-up with your personal trainer planned in less than an hour. That might not be the right time to start watching a two-hour long Top Gear special. Rather, you’d probably enjoy watching a quick 10-minute TED talk. You’re leaving on holiday to Mexico in a few days? Why not check out this documentary on The History of the Maya.
However, day of week, time of day, device and location are excellent places to start off from. Not the least as this will make recommendations match our ‘routines’ even closer. Having some routine in our lives benefits us – it makes most of us feel comfortable, and that remains what Netflix actually sells to us; not TV shows, but some enjoyable moments of relaxing, and not having to think too much. Perfect recommendations lay at the bottom of that.
Curious to learn about ‘how recommendations are made’? Then join me in this Introduction of Recommender Systems course on Coursera, starting on September 3rd. To find out more about context and how contextualization helps deliver (way) better recommendations and an improved user experience, download and read the free Argus Labs whitepaper ‘Putting Music in Context’.