The secrets of Netflix’s recommendation system - and why it may not work for you

Yet, you’re not alone if you feel like Netflix doesn’t quite “get” you.

“When I started at Netflix 12 years ago, we were learning how to crawl with regards to personalisation”, says Todd Yellin, Netflix’s vice president of product. “Now, I would say we’re in our adolescence. We’re still not perfect – we’re far from perfect. I think we’re good. I strive for great.”

But how do recommendations actually work? And where do the flaws lie? Check out our handy layman’s guide below.

What's the theory behind Netflix’s Recommendation system?

There are two main ideas at play here – and they both come from what Netflix has learned by surveying user data over the years.

Firstly, they know that most of their users don’t want to waste too much time looking for something to watch.

“The typical person isn’t going to look at thousands of titles, they’re gonna look at an average of 40-50 titles on every given session,” says Yellin.

Netflix thus have a small window in which to pique your interest, or risk losing your attention – so their primary focus is to make sure that the first things you see when you log on are titles that you want to watch.

Secondly, they have learned along the way that what users say about how they use the service and their actual behaviour do not always correlate.

“A lot of people tell us they often watch foreign movies or documentaries. But in practice, that doesn’t happen very much”, said Carlos Gomez-Uribe, Netflix’s former vice president of product innovation in an interview with Wired in 2013.

Similarly, they know that you may choose to rate a smart documentary that you’ve watched once with 5 stars, while you may give a lower rating, or no rating at all, to the Adam Sandler movie you’ve watched four times this year. This is, presumably, one of two reasons why they decided to remove the star-rating system in favour of a thumbs-up, thumbs-down model. More on the second reason later.

But HOW does it work?

Put simply: data.

A number of lucky Netflix employees are paid to watch all of the titles and mark down any number of defining elements that occur. For example, a movie such as Wall-E is tagged as follows: Warm-spirited, sparse dialogue, satirical, and so on. There can be any number of tags – the more the better.

Then the algorithms comes into play. The more you watch Netflix, the better it aims to understand your tastes by compiling a profile based on recurring tags in the shows you are watching.

So, if you’ve watched Marvel’s Jessica Jones, which may be tagged as dark, with a strong female lead among other things, it is quite likely that Orange Is the New Black will come to the top of your deck.

Every category on your front page is personalised based on your viewing behaviour, pushing content that matches the patterns you have been unknowingly drawing out, to the front. The algorithms also take into account specific info about the user - what kind of device you watch on, and what times you tend to watch.

If you're interested in finding out more, Yellin made a handy explainer video – check it out below.

Why am I still getting recommendations that I have no interest in?

This is likely because Netflix have taken a decidedly hard line with regards to the subjectivity of taste.

“When personalisation is at its best – it’s not really about ‘well, that’s bad, that’s good’” says Yellin, “it’s about whether that’s bad for this person, if that’s good for this person.”

This line of thinking may have come into play in the decision to remove the 5-star rating system in favour of thumbs up, thumbs down. It is no longer possible to determine how fellow Netflix users feel about a show – thumb scores are not visible but go toward a “match” rating, which signifies the likelihood that you will enjoy a title based on the aforementioned algorithms.

It’s no coincidence that this means million-pound Netflix projects are no longer at risk of being labelled with a poor star-rating for all to see.

The new system allows for shows that have only been liked by a relatively small number of people to be recommended to you purely on the basis that they bear some of the same tags as shows you like.

Taste is subjective, but with a show such as Gypsy, a Netflix Original starring Naomi Watts, which premiered to dire reviews in June, and was cancelled within two months of its debut, the probability that the average user will like it is significantly lower than with a show like House of Cards, their flagship drama that was well received across the board, even if it contains similar elements according to the tagging system.

“I don’t like the word cancel, because Gypsy is going to be on the service for many years to come. We just decided not to do another season. And Gypsy will still be personalised to many people for years to come,” says Yellin.

Is it likely to improve in the future?

There is no reason to doubt Netflix’s ability to innovate and improve their service. The recommendations are steadily improving – a "Bad Netflix Reccomendations" Twitter account burned out in 2015, having seemingly run out of good material.

And Netflix’s refusal to accept a general consensus on the quality of shows may become less of a problem in the future, as their algorithms and whatever other personalisation technology they have in the works improve.

And, if Elon Musk is to be believed, artificial intelligence could be capable of anything. But even if there is a computer uprising, at least your MacBook will have a more nuanced understanding of your relationship with Wes Anderson’s oeuvre.

Last updated 8th September