More than 80 per cent of the TV shows and movies people watch on Netflix are discovered through the platform’s recommendation system. That means when you think you are choosing what to watch on Netflix you are basically choosing from a number of decisions made by an algorithm.
Netflix offers thousands of TV shows available for streaming. It recommends titles for each user. If you use Netflix you may have noticed they create really precises genres: Romantic Dramas Where The Main Character is Left Handed. How do they come up with those genres? How to they deal with giving great recommendations to their 100 million-plus subscribers who are already used to getting recommendations from pretty much every platform they use? Machine learning, algorithms and creativity. Those are their magic tricks to help break viewers’ preconceived notions and find shows that they might not have initially thought of watching. For those who are still trying to figure out what an algorithm is, it’s basically a set of databased instructions that tells Netflix what to do.
The recommendation system works putting together data collected from different places. Recommended rows are tailored to your viewing habits. That’s why you can tell when your little cousins have been using your account to watch a billion hours of Peppa Pig. In this case, algorithms are often used to facilitate machine learning. Systems like Netflix based on machine learning rewrite themselves as they learn from their own users. Every time you press play and spend some time watching a TV show or a movie, Netflix is collecting data that informs the algorithm and refreshes it. The more you watch the more up to date the algorithm is.
The collected data is multi-faceted and complex, but it involves way more than just processing the genre of a program a user is watching and recommend him or her dramas, romances or comedies. Todd Yellin, Netflix’s VP of product innovation, told Wired in 2017: “what we see from those profiles is the following kinds of data — what people watch, what they watch after, what they watch before, what they watched a year ago, what they’ve watched recently and what time of day”. The Netflix experience is driven by a number of machine learning algorithms: ranking, search, similarity, ratings and more. They can’t offer their entire catalogue at once so they must curate it. As quality and taste are rarely the same thing, Netflix cannot work as Rotten Tomatoes, Pitchfork or IMDb, they have to know their users and get recommendations tailored to each individual.
Netflix works with taste groups. Each viewer fits into multiple groups and these affect what recommendations pop up to the top of every onscreen interface, which genre rows are displayed and how each row is organized. If your viewing patterns are similar to another user Netflix will serve up recommendations based on the behavior of that other users as well.
The tags that are used for the machine learning algorithms are the same across the globe. Netflix has hired real life humans to categorize every bit of TV shows and movies and apply tags to each of them in order to createhyperspecific micro genres such as “Visually-striking nostalgic dramas” or “Understated romantic road trip movies”.
Each of these data factors comes together to identify which taste group you fit into. Each user’s screen gets populated — left, right, and top to bottom — is based on which groups they belong to.
Why Rows Anyway?
Chris Alvino, Machine Learning Engineer at Netflix, explains they choose rows to make it easier for members to navigate through a large portion of their catalog. By presenting coherent groups of videos in a row, providing a meaningful name for each row, and presenting rows in a useful order, members can quickly decide whether a whole set of videos in a row is likely to contain something that they are interested in watching at that precise moment. This allows members to either dive deeper and look for more videos in the theme or to skip them and look at another row.
Each device has different hardware capabilities that can limit the number of rows displayed at any one time and how big the whole page can be this is why Netflix must be aware of the constraints of every device.
Each row can offer a unique and personalized slice of the catalog for a member to navigate. Part of Netflix’s challenge is to create useful groupings of videos in order to highlight the depth in the catalog and help members not only reinforce their areas of interest but also find new ones. Recommendations ought to be fresh and responsive but also stable so that people are familiar with their homepage and can easily find videos they’ve been recommended in the recent past.
An Image is Worth a Thousand Words
Netflix has implemented recently a new recommendation algorithm based on artwork. It serves up unique tailor-made images to its subscribers. These images are specially designed to keep you stuck in Netflix. It takes into account a lot of the same data factors we’ve mentioned.
Gopal Krishnan explained all about this new algorithm on his technical blog post. Netflix has been working to create a framework that allows them to effectively intersect big data with creativity helping users discover shows and movies they’ll enjoy faster and prevent them of being overwhelmed with Netflix’s HUGE catalogue. As a result of that investigation, they now have the unique ability to understand which images work best for each user.
They say that if they don’t capture a user’s attention within 90 seconds, he or she will likely lose interest and move onto another activity. Having such a short time to capture interest, images becomes the most efficient and compelling way to make users discover the perfect title as quickly as possible.
They have built a system that tests a set of images for many titles on their catalogue helping display a compelling image to drive engagement. Through many experiments and tests, Netflix arrived to the conclusion that seeing a certain range of emotions actually compels people to watch a TV show or movie. This is likely due to the fact that complex emotions convey a wealth of information to users regarding the tone or feel of the content, but it is interesting to see how much members actually respond this way in testing. An example of this is seen in the recent winning image (“winning” means it drove the most engagement) for the second season of Unbreakable Kimmy Schmidtbelow:
That winning image is the one that would work best for a majority of Netflix users. But they have pushed even further and given the enormous diversity in taste and preferences, they decided to put together different artwork for each user to highlight the aspects of a show or movie that are relevant to them.
In order to choose which image each user has on their feed, Netflix focuses on what other shows and movies users have been watching. For example, a member who watches many movies featuring Uma Thurman would likely respond positively to the artwork for Pulp Fiction that contains Uma. Meanwhile, a fan of John Travolta may be more interested in watching Pulp Fiction if the artwork features John.
Not all the scenarios for personalizing artwork are this clear and obvious but data does all the work in order to choose the artwork for each user and improve the Netflix experience (and keep users binging, of course).
Personalized page generation is a challenging problem that involves balancing a multitude of factors and probably this is just the beginning. There is a lot of potential to improve the homepages and several dimensions in which to Netflix will certainly experiment.
As regarding technology, way to go! But, will the quality of TV shows and movies lower just to please more users and keep them binge watching? Will we find a new favorite show every week? What is binge watching doing to our brains? Watching episode after episode of a show feels good — but why is that? Maybe it’s all about reaching a balance. We don’t know yet but it’s certainly a really interesting issue.