How Netflix's new image algorithm works
As a Netflix subscriber, you are constantly participating in a range of behavioral experiments to produce data that allows the company to recommend movies and TV shows it thinks you'll like. What you watch next, what you give thumbs up and thumbs down to, when you quit watching a not-so-bingeable show: All of this information can help engineers create a recommendation algorithm.
Historically, the company's tech team has tested algorithms on designated batches of users to see which leads to more watch time, more titles watched -- anything that keeps viewers on Netflix. Once they figure out which algorithm performs best, they roll it out to everyone, but all that time spent testing means most of the people on the platform aren't getting what the company terms "the better experience."
The new image algorithm, though, works in real time to project the image it thinks you'll respond to, and continues collecting data to improve its performance. And it's doing the same for 100 million other subscribers, collecting THAT information for further customization.
In the end, you get an image that Netflix's algorithm thinks will entice you to watch, but the key is that the picture could change tomorrow as the system "learns" more about you and subscribers like you. For any given show, there might be a dozen possible images loaded, which are ranked according to "context." So, in the Good Will Hunting example, the Robin Williams image might be ranked fifth-best for a user Netflix determines is romance-centric, but first for a comedy profile. For each title, you'll get the image with the highest rank based on your profile.
It's not a revolutionary or entirely unique approach to contextualize each user's experience, but doing so for such a huge number of people at on-demand speeds could open the door to further customized aspects of the platform. In their post, the engineers say they could potentially apply the same method to the short descriptions of each title, or even trailers -- if you don't like love stories, a description for an action-romance might focus more on action, though the Netflix engineers are conscious of straying too far from truth, which leads to what they call "clickbait images." There's no point in hyping up a movie a user is unlikely to enjoy, because it risks alienating and disengaging that user.