Contents
Online cinemas experienced a real boom in 2020. But if they hardly set up a pandemic, then they learned to manage the audience 100%. Explore how Netflix’s algorithms manipulate our choices
Step 1: personal recommendations
Today’s recommendation system is based on Netflix’s own research and development in AI and machine learning. The algorithm has been perfected over the years, training it on a growing database of subscribers. But by 2016, 80% of users were choosing what to watch based on recommendations.
A year later, the head of Netflix, Reed Hastings, said that their main competitor is not other streaming services, but sleep.
The collection of big data is the main reason why Netflix strives to keep users as long as possible: the more data, the more accurate the algorithms work. Moreover, the data is collected not only on the basis of the history of views and likes. For example, when watching the first interactive film Black Mirror. Bandersnatch, users had to choose which path the story would take next. Data about who pressed which buttons was also used to train the neural network.
Netflix makes personalized recommendations for you and places them on the home page. To do this, neural networks use two ways:
One of the recent Netflix personal recommendation scandals has taken place in the United States. The service was accused of showing teenagers movies and series about suicide, and this leads to sad consequences. One such series is 13 Reasons Why, which led the family of a teenager who killed himself to sue Netflix. However, representatives of the online service insist that any restriction of recommendation algorithms is akin to censorship: it’s like restricting the news feed. Netflix also emphasized that they do not collect personal data about users, including age.
How does it work in practice? The system endlessly offers you different selections and remembers what worked in the end: that is, which of the recommended films you started watching or even liked, and which ones you abandoned in the very first minutes. Then he uses this data to rebuild the recommendations feed, guessing more and more accurately over and over again. And so – ad infinitum.
In addition, the system notes what time of day you usually watch, in what region and from what devices – and ignores gender and age. All this is processed using special algorithms, and as a result you see several series of recommendations, for example: “Watching now”, “Trending”, “Award-winning comedies”, and so on. The higher the row, the more, from the point of view of the system, these films suit you.
In April 2021, the Play Something feature was added to the usual recommendations. It allows you to play random movies and series if you can’t choose what to watch. This selection is also compiled by the algorithm based on your preferences and preferences of similar users. So the system seeks to keep you on the platform for as long as possible.
Step 2: personal posters
The first thing you see when you flip through a catalog of movies and TV shows is the poster, which is also the cover. You don’t know anything about the movie yet, haven’t watched the trailer yet and haven’t had time to read the description, and the service has literally a couple of seconds to hook you. How to do it?
To select the best covers, Netflix uses AVA – Aesthetic Visual Analysis, “analysis of visual aesthetics.” Each frame is assigned different attributes with a specific weight in importance, for example: contrast, brightness, number of faces, skin tones, presence of a naked body, motion blur, composition symmetry.
Comparing all these signs, the algorithm selects the most successful frames for the poster and even edits them.
The algorithm then analyzes which images are most frequently clicked on by you or users with similar preferences. Then he selects pairs from a thousand frames and shows them to you one by one – this is the so-called A / B testing. At the same time, they also test posters proposed by different versions of the algorithm – old and new: this is how the service finds out which one works better.
Based on your previous selection, the service arranges the entire feed so that you see posters of a certain type and click on them.
For example, if you like comedies, then as a poster for “Good Will Hunting” you will be shown a frame with a smiling Robin Williams. And if you prefer romantic dramas, you will see the main character with his girlfriend:
The neural network takes into account which country you are in. Compare two posters for “The Eighth Sense” for residents of Germany and the USA:
This is how the algorithm understands that you pay attention to certain actors:
Step 3: Personal Trailers
As with posters, special algorithms for continuous online learning are responsible for trailers – “contextual bandits”. They apply context—that is, everything they know about you—at every step of the selection process.
In the same way as with posters, AI based on machine learning selects personalized trailers for different audiences, taking into account preferences, gender, age, country, and even time of day. And even more than that: Netflix plans to use neural networks to select the necessary frames and mount thousands of trailers from them for different users.
For example, for fans of romantic films, the algorithm will offer cutting frames or a separate scene with a couple in love – even if it is an action movie in which romantic scenes occupy 0,5% of the entire film. If you often watch films with the same actor, then you will be shown a trailer with him.
Trailers are just the beginning. Hollywood uses neural networks to predict which movie will be a box office success and invest in it. And they even make movies using a neural network.
What algorithms does KinoPoisk use?
The Russian streaming platform is part of Yandex and uses its algorithms. Like Netflix, KinoPoisk makes personal recommendations for you: based on your viewing history or based on similar users, as well as movie data: release date, genre, awards, ratings. At the same time, the system also takes into account the search history of this user on Yandex in order to more accurately determine his interests.
But even before the advent of online cinema, ratings were the main feature of the service. Like on IMDb, Rotten Tomatoes or Afisha, users give ratings to films and series. Users noted that not all votes have the same effect on the rating.
The service has been repeatedly accused of cheating voices for some domestic films that failed at the box office. For example, in 2016, users suspected that the ratings of the films “Earthquake” and “Mafia: A Game of Survival”, and later “Crimea”, were cheated. In all cases, the rating soared immediately after the release and grew in a matter of hours, which is not at all typical even for super-popular blockbusters. KinoPoisk claimed that the accounts from which positive ratings were given were hacked.
The opposite happens: the film “Crimean Bridge” immediately after the release received one units, which from the outside also looked like an attack of fake accounts.
Now, according to the developers, the service uses a complex system to protect against cheating. In particular, it recognizes suspicious ratings and does not take them into account in the overall rating of the film. Dropout criteria are constantly updated, following new ways to bypass the algorithm. Also, for the period of film distribution, the service publishes only reviews from verified users: these are active visitors to KinoPoisk, who are evaluated according to a number of criteria in the history of interaction.
Step 4: Packaging Videos for Different Devices
Netflix takes into account not only what you watch, but also where. Depending on your region, device subscription plan, or browser, you’ll see videos in a specific format. In total, the service supports more than 2000 different systems: TVs with Smart TV, set-top boxes, smartphones, tablets, computers and laptops with all existing operating systems.
But it’s not just about the quality, format and resolution of the file. When a video is uploaded to the platform, Netflix’s smart algorithm checks it for dead pixels and other defects. Then it splits into several different files: with different audio tracks and subtitles, resolution, video and sound formats, as well as those optimized for different Internet speeds. For example, from the sources of the second season of Stranger Things, the output turned out to be 9 files: it took 570 hours of processor work. And all this to ensure you get the best viewing experience on any device, even with poor internet.
Step 5: Launch Movies and Series in Different Countries
The service also uses artificial intelligence to promote new projects. Netflix’s algorithms use audience data in each country to calculate which movies or shows to release where and when, how to best present them, and which other movies to link to. If the system sees, for example, that a certain drama will do better in Spain, it will help marketers and producers direct more resources there and prepare the Spanish titles and dubs first, and then the rest.