How Netflix knows what you want to watch before you do

It’s the age-old question we all ask ourselves after finishing any book or television series we’ve committed hours, days and sometimes months of our lives to…

"What now?"

If you, like millions of other Australians, have signed up to Netflix, the wildly popular streaming service answers this question for you. Dozens of TV shows and movies vie for your attention on the Netflix homepage, and they're not placed there all willy-nilly. Netflix invests a lot of time and effort into curating those titles specifically for you, crafting recommendations based on your tastes such that your homepage will almost certainly look completely different to the home page of your friends, family or anyone else.

While these recommendations are far from perfect (no Netflix, I do not want to watch Pixels!), they tend to hit far more often than they miss. More importantly, they're constantly evolving, growing more accurate with every additional show or movie you watch on the streaming service.

So how does Netflix do this? No, there aren't a million poor desk jockeys poring over your viewing habits and trying to match you up like some archaic dating machine – Netflix's magic is all thanks to algorithms. Algorithms are, essentially, complex rules that automatically process data to reach certain outcomes. In Netflix's case, its recommendation algorithm looks at how you've engaged with the service and scours the Netflix servers to pick out new movies and TV shows it believes you'll like.

Figuring out what you'll like isn't easy – how many times have you hemmed and hawed over what to have for dinner? To accomplish this, the recommendation algorithm sifts through an enormous amount of data, analysing not just the shows and movies you've watched, but when you've watched them, what you've watched before and after them and whether you've binged them or spread them out over multiple sessions. This data is then combined with a comprehensive catalogue of "metadata" which categorises the shows and movies based on an extensive array of content tags. These tags can be as simple as the genre of a show or the lead cast, or they can be as in-depth as the type of music featured on the soundtrack or the philosophical themes tackled by the plot.

By combining this data together, the recommendation algorithm is able to link shows like Jessica Jones with Orange Is The New Black based on their strong female protagonists, as well as find common ground between the dangers of technology in Black Mirror and the grim societal prejudices exposed in Luke Cage.

How did this algorithm come about?

Here in Australia, we received Netflix Australia in its prime, with 4K streaming, partitioned profile features and optimised algorithms. With all these fancy features, it’s hard to believe that Netflix started as a simple DVD delivery service. Even harder to believe is that Netflix's recommendation algorithm was in place way back when people were still receiving big red envelopes on their doorstep.

Cinematch was the first algorithm Netflix used to predict movies and TV shows that users would enjoy. It was so good at doing its job that 50 percent of Netflix users mailed their recommended film back with a five-star rating.

Netflix identified recommendations as an integral part to its operation early on, and focused on continually improving the algorithm. By 2006 the company had gained traction, and the desire to improve the algorithm increased. It was around this time the company launched a competition to find an algorithm that could outperform Cinematch.

The "Netflix Prize" looked to award $1 million to an individual or team that could match the accuracy goals for suggesting movies based off users’ personal preferences. Three years later, the two teams "Pragmatic Chaos" and "The Ensemble" submitted their entries within 14 minutes of each other, with Pragmatic Chaos beating the competition by the skin of their teeth. Both competitors managed to improve upon Cinematch’s algorithm by about 10 percent – a big success in the eyes of Netflix.

Today, the algorithm looks at everything the original Cinematch did with a few added tweaks, including observing user behaviour, social trends and popularity divined from its ever-evolving user-rating system.

What's next?

Because its recommendation algorithm is based in software, Netflix is able to constantly improve how it decides what you might enjoy. Not only does the algorithm hone its relevance the more shows and movies you stream, it also learns from the millions of hours everyone else in the world pours into the service. Trends and patterns emerge from the wealth of viewing data Netflix has at its disposal, and this data then informs the recommendation algorithm on what kind of recommendations are working best. Better recommendations mean more streaming, which means more data – and so the whole cycle repeats itself.

That's not the only way Netflix is improving its recommendations, either. Just recently, it rolled out a new "artwork personalisation" system to better highlight why it is recommending a particular title to you. This works by customising the thumbnail artwork of a particular title based on the rules that drove the algorithm to pick it.

For example, if the algorithm decided that your history of watching action movies featuring breakneck car chases makes The Bourne Legacy a likely candidate, it might use an image of one of the movie's highway scenes to entice you to check it out. On the other hand, if it had reached the same recommendation based on your affinity for political thrillers, the image might instead show off a clandestine meeting at a cafe.

This is just one of the many techniques Netflix employs to constantly improve its recommendations, and it certainly won't be the last. Even if you're not particularly impressed by the titles the service is currently suggesting for you, just keep watching; in a year, you might find that Netflix knows your tastes even better than you do.

If you don't believe so maybe instead check out new Netflix titles or use our Netflix random generator.