Why YouTube recommends the videos it recommends

How?Social Networks show us different types of content is still a mystery to many people, who even think that these platforms can read their minds.?However, beyond psychic powers, these predictions and recommendations are based on our tastes and the type of content we regularly see.

Now, it is?Google?that has decided to also shed some light on the mystery of?YouTube video recommendations.?

What is the YouTube recommendation system

The recommendations are nothing more than the different videos that we find on the home page when entering the platform, or the videos that appear in the "Next" panel when we are viewing a video.?That is, they are the videos that YouTube shows to?help users find content of interest without having to search for it.

Currently the content recommendation?is mainly based on the tastes of the users and their viewing history,?but this has not always been the case.?What's more, YouTube's recommendation system has evolved extensively since its implementation?in 2008, when it was simply based on the popularity of videos?.?Thus, if cooking videos were the most viewed at that time, they would appear directly on the user's home page even if they had not shown interest in this type of content.

From there, the system improved: between 2011 and 2012 the platform incorporated classifications to identify?racist or violent videos and avoid their recommendation,?and began to take into account the?viewing time?as a recommendation factor, helping to improve the quality of the system. .?In the same way, between 2015 and 2017, YouTube downgraded?tabloid content?to avoid its recommendation, and reduce misinformation.

On the other hand, it also began to implement a?machine learning system?to help deploy content in sensitive groups, such as minor users, a predictive system was also incorporated to?avoid the recommendation of videos that may present risks for them, such as challenges.?All this finally gave way to the recommendation system that the platform currently has.

How YouTube video recommendation currently works

The idea of the recommendations is to provide users with a list of content that is of interest to them or that has some associated value for them.?To achieve this, the video social network?takes into account the viewing habits of users and compares them with those of other users with similar habits.?Thus, the system takes this information to suggest content that may be of interest to users.?For example, if we like videos on digital marketing and the system notices that the majority of users who watch the same type of videos also watch videos on technological news, it is very likely that it will begin to recommend these videos to us.

However,?privacy?is also an important factor to guarantee a better browsing experience for users.?For this reason, YouTube also has functions that allow you to control the access that the platform has to the information of the viewing history.?In this way, users can?delete content from their history, clean the history completely or even pause their monitoring?, whether it is search or viewing.

How recommendations are personalized on YouTube

Currently, YouTube's recommendation system is constantly evolving,?learning daily through the analysis of 80,000 million parameters or, as Google calls them, “signals?”.?These signals complement each other and form the different aspects that are taken into consideration when personalizing the recommendations, denoting the tastes of the users.

The main signals that YouTube takes into account are?clicks, watch time, survey responses, sharing, likes and dislikes,?although none of them is, individually, decisive when it comes to recommending a video.?The platform complements the information from all of them to make the most appropriate decision.

Clicks

The click is the most basic aspect of interaction, as it implies that the user has an interest in the content of the video.?However, as Google reminds us,?“clicking on a video does not mean that the user has actually seen it.?This is why, in 2012, we started to look at playtime. '

Display time

Playback time?is taken into account here?, that is, what videos users have watched and for how long.?This signal makes it possible to differentiate the specific content that draws the most attention from users, and thus offers a more assertive recommendation for the user.

In the words of Google,??If a tennis fan watched 20 minutes of clips of Wimbledon highlights and just a few seconds of match analysis videos, we can safely assume that they were more interested in watching those highlights.?When we first incorporated watch time into recommendations, we immediately saw a 20% drop in views, but we found it more important to deliver a better viewer experience. "

Sharing, likes and dislikes

Other signals to determine which videos to show users are?their reactions?.?Therefore, if a user has shared or liked a video, they will likely want to see other videos of this type.?From this, the system predicts the probability that the user shares or likes other content, and based on this it recommends other videos.?"If you don't like a video, it's a sign that you probably didn't enjoy watching it,"?Google explains.

Survey responses - valuable watch time

“To make sure that viewers are satisfied with the content they watch, we measure what we call?'valuable viewing time'?, that is,?the time a user spends watching a video that seems valid to them.?Valuable watch time is measured through surveys, in which the user is asked to rate the video they have watched from one to five stars.?This gives us a metric to determine how satisfied the viewer was with the content. "

In these surveys, YouTube also asks?the reason for the rating.?But these surveys are not always answered, so a machine learning system is incorporated that adds?estimated responses?, based on those already carried out.

The human factor in the fight against misinformation

Most of the content consumed on YouTube is for?entertainment?, such as music videos, jokes, gameplays, etc.?And for this type of content the signals we mention below work perfectly.?However, with the growing reach of digital media many people come to watch?news or informational videos.?Issues in which the quality and veracity of the information are very important, and therefore these signals can be seen as a more subjective factor for their evaluation.?"Someone can report that they are very satisfied with the videos that claim that the Earth is flat, but that does not mean that we want to recommend this type of low-quality content,"?explains Google.

In this way, YouTube has been reinforcing its commitment to a more responsible recommendation in these types of informative content.?With which it joins the?fight of digital platforms against disinformation,?a problem that strongly affects the reputation and credibility of social networks.

'We can exercise this control by using classifiers to identify whether a video is' solvent' or 'borderline'.?These classifications?are made by people?who evaluate the quality of the information on each channel or video.?These evaluators are recruited around the world and receive detailed, publicly available grading guidelines in their training.?We also rely on?certified experts?, such as doctors, when the content is related to health.

To determine the creditworthiness of content, evaluators?answer a few key questions?: Does the content deliver what it promises or does it achieve its goal??What kind of experience is needed to achieve the goal of the video??What is the reputation of the person speaking in the video and the channel they are on??What is the main topic of the video (eg news, sports, history, science, etc.)??Is it primarily a satire??These answers and others determine the creditworthiness of a video.?The higher the score, the more the video will be promoted among news and information content.

In determining whether content is questionable, evaluators take into account, in addition to other factors, whether the content is inaccurate, misleading, or misleading;?if you show intolerance or lack of sensitivity;?if it is harmful or could be.?The results are combined to obtain a score that indicates the?probability that a video contains harmful or questionable misinformation.?Any video classified as doubtful is relegated by the recommendation system.

YouTube also clarifies that?dubious, offensive or sensationalist content does not generate better results for the platform?, compared to what might commonly be thought:?“when we began to relegate explicit or sensationalist content, we observed that?the viewing time increased by 0, 5%?over the course of two and a half months compared to when we didn't set any limits. "