Here’s How You Found "Stranger Things": Algorithms in Netflix Recommendations

By Ore James

Image Credit: Complex Magazine

With most teens stuck at home in the midst of a pandemic, Netflix and other streaming services have become somewhat of a refuge. As browser extensions like Netflix Party facilitate remote interactions with friends, the relative normalcy provided by streaming may explain its popularity among the quarantined - of course, the data science behind these recommendations also plays a role in keeping us glued to the platform.

Most familiar with Netflix have probably seen genres like “critically-acclaimed movies about friendship” or “comedies for hopeless romantics” mixed into their homepages. However odd these micro-genres may seem, there exists a solid method behind them - personalized recommendations rely on machine learning algorithms to keep subscribers engaged, aiming to prolong our binge-watching sessions.

Recommendation systems are simply platforms to suggest content based on existing user preferences. The Netflix recommendation system compresses its large streaming library into personalized, easily-navigable rows using machine learning. 

Within machine learning, systems regularly rewrite their algorithms - or data-based instructions - according to user data. Essentially, the systems collect data from users, learn from the data, and apply what it learned to make decisions. As 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.” In addition to this data, Netflix relies on man-made tags, which categorize the service’s content, to determine the types of media users prefer. Taking all of this into account, machine learning algorithms interpret these data sets to ultimately decide which content to recommend.

Machine learning can take on highly specific forms to maximize user experiences. Recently, for example, Netflix incorporated an artwork-based algorithm to further personalize recommendations. The algorithm uses user preferences to determine which artwork will appear next to movies and shows. Machine learning comes into play when the algorithm chooses the art - data suggesting a user likes horror movies, for instance, may compel the algorithm to choose dark and chilling artwork for “Stranger Things.” By incorporating artwork, the algorithm demonstrates data can combine with a sense of creativity to further increase user engagement.

Image Credit: UX Planet

The downside to the Netflix approach of highly specific suggestions, as many subscribers have observed, is the fact that data from a customer's watch history often fails to reveal their actual tastes. For instance, within the current system, users who accidentally click on a documentary may face homepages cluttered with docudramas they have no interest in watching. The potential for users to miss out on content - or, conversely, face a homepage of content they don’t want to see, presents a glaring flaw in recommendation systems that can ultimately harm the subscriber experience. Content creators have voiced similar concerns - when Netflix canceled sitcom “Luca and Bertie” after one positively reviewed season in 2019, show creator Lisa Hanawalt pointed to the algorithms as a cause for low viewership. 

Of course, machine learning is just one subset of artificial intelligence. Deep learning is a sector of machine learning in which a machine uses artificial neural networks, inspired by the brain’s neural networks, to “train” itself to make more accurate predictions. In practice, deep learning enables artificial intelligence to “think” and learn. While the technology is more commonly used to identify photos or audio, services like employ deep learning to recommend movies by adapting to user preferences in real-time, aiming for more accurate movie recommendations. Netflix itself seems to be slowly moving away from strict machine learning, following competitors like HBO Max; in August 2019, the service began beta-testing a Collections section, which relied on humans, not algorithms, to group titles for users. 

While Netflix’s complex algorithms currently sort its thousands of titles well enough to keep many of us on the platform, it’s clear the future possibilities of machine learning, deep learning, and creativity in streaming are endless.