Presentation: Artwork Personalization @Netflix

I lead one of the Machine Learning and Recommendation teams at Netflix. We're responsible for the end-to-end machine learning that decides what shows up on the Netflix homepage across all our different experiences. When you log into Netflix, my team is responsible for what rows of TV shows and movies you see on the homepage. We select rows and titles to help our members quickly find something great to watch. For example, whether you see Popular on Netflix or Continue Watching at the top of your homepage. We also work on algorithms to personalize the imagery that we display for each title across the experience.

There are over 130 million Netflix subscribers and each home page has many different recommendations on it. Over 80% of what people watch comes from our recommendation systems. Each time you log in, you're typically seeing something new. So we’re serving up a huge number of recommendations to people on a regular basis.

Our personalization don't end with selecting what title we want to show to the member.  Each section of the screen is a precious piece of real estate so we want to ensure the member understands why this particular title is being shown to them, especially for a new original series or movies that we create. A great new title may only have one chance to really resonate with somebody, so we want to give it the best opportunity to make the connection with them.

To do this, we select the imagery based on the person’s viewing habits. For example, if we take a romantic comedy, maybe you've watched a lot of comedies so we show you a funny image from that movie. If another person watches more romances we may use an image of a couple from the movie on a date. It’s about helping a member understand what's in it for them by presenting the most appropriate imagery we have. This helps them find and enjoy that film.

The target audience for the talk is people who have some background or are interested in machine learning. They should know some basic machine learning but would like to learn more about how to use bandits and contextual bandits. Bandits add new technical challenges by having to close a feedback loop. It's not just a classification or regression problem. You're actually putting something out there and then needing to pull the data back. Once it comes back, you then update your models and put them out again. It’s a continuous loop of improvement. So it’s an interesting application of machine learning that should be of useful to people facing similar challenges.

Making sure that when we're recommending titles we're actually showing the ones people want to watch and will enjoy. We know that when we show titles on the homepage, people are likely to click on them and watch them because they're ranked highly. But we want to make sure that our recommendations showcase the best way for people to spend their time on Netflix. We don’t want our recommendations to be influenced by feedback loops just due to us ranking them highly in the past because it may mean we don’t end up surfacing the content that people will enjoy the most.