It is quite hard for users to come out of YouTube watching just one video. That is the power of YouTube’s recommendation engine which suggests what users should watch in the first place and they what they should watch next.
“Today, Recommendations drive a significant amount of the overall viewership on YouTube, even more than channel subscriptions or search. And, we are thinking about it in a responsible way. Our goal is to help connect viewers to high-quality information by minimising the chances of them seeing problematic content, explained Cristos Goodrow, VP Engineering, YouTube in a recent call. He said YouTube’s goal was to have views of borderline content from recommendations kept below 0.5% of overall views on YouTube.
So how does, the recommendation engine on the world’s most popular video platform work?
Recommendations were originally built on the simple principle that it should help people find the videos they want to watch and give them value. Users now encounter recommendations at two places — one on the homepage as they enter YouTube which is a mix of content based on past viewing and subscriptions as well as latest news. Then they also see recommendations in the “Up Next” panel as they are watching a video.
YouTube tries to predict what a user would like to see next based on what they usually like to watch, based on their own preferences and interests. It does not use connections from the social network to recommend what to watch next.
YouTube takes signals from a mix of user behaviours which are strong indicators that they like a video. So clicks on videos, watchtime and shares are taken as good cues. Along with these, YouTube also measures “valued watchtime” using user surveys that ask them to rate a watched video on a scale of one to five.
Yes, when users delete their watch history they also disable personalised recommendations in he process. Removing a certain video from watch history also pauses recommendations linked to that video.
What is borderline content and how does YouTube keep it off recommendations?
Since 2011, YouTube has limited low-quality content from being widely viewed via recommendations. In 2015, it started taking steps to demote “sensationalistic tabloid content” on the homepage. Within a year, YouTube was able to predict “the likelihood of a video to include minors in risky situations” and started removing those from recommendations.
In recent years, YouTube has been keeping “problematic misinformation and borderline content —that is content that comes close to, but doesn’t quite violate our Community Guidelines” out of recommendations. The platform does this by using classifiers to identify whether a video is “authoritative” or “borderline” with the help of human evaluators placed across the world.
How does YouTube factor in languages while recommending videos?
Interestingly, Goodrow told The Indian Express that “the recommendation system in general, doesn’t have a notion of language”. He elaborated: “Most of the recommendations start or are based on what we call code, watching videos that get watched again. This tends to ensure that that videos of the same language, get recommended together. Or if people speak only one language they mostly only get videos in that language.” This means there’s nothing in particular that’s done to sort of filter or determine what are the languages that a particular user speaks with respect to recommendations. “It’s really due to just the frequency of two videos being watched together.”
How does YouTube recommend videos when a user is in incognito mode?
Goodrow says in such a situation they “usually don’t have much information to go on”. In such a case they base recommendations based on the devices and what is popular in your location. “At that time we try to ensure that there is some diversity.”
How does YouTube handle recommendations for search queries?
When users are searching for a news event, YouTube uses the top shelf to showcase news from authoritative and trustworthy sources. This also means content from untrustworthy sources are demoted and further down in the results.
Newsletter | Click to get the day’s best explainers in your inbox