Image Credit: Interior Design Magazines
By Trisha Sengupta
From years of wasting away my life on YouTube, I have often heard creators complain about the “YouTube Algorithm” and how it damages their career by demonetizing them or not recommending them. But what is the YouTube Algorithm and how does it work? Is it even an algorithm? By exploring YouTube and the mechanisms by which YouTube recommends videos, these questions and more can be answered.
The YouTube Algorithm has evolved over the years. Before 2012, it focused simply on view count; videos with more views would be recommended to more viewers. However, this lead to the problem of clickbait where creators added purposefully catchy titles without actual substance in their videos. And so, YouTube changed its algorithm to account for view duration, or watch time and time spent on the platform or session time. This caused creators to delay the time taken to deliver on promises that the video’s title makes. The algorithm’s changes also led to creators being obliged to make high quality videos while increasing the rate at which they were produced. People could not make high quality, lengthy videos. It also explains why so many popular YouTubers at this time were gamers as they could produce long videos in short periods of time without a lot of editing.
From 2016 onward, YouTube changed its algorithm again, releasing a lengthy paper describing how the new process works. In their new system, YouTube employs deep learning to improve their recommendation process. YouTube is a platform with 300 hours of content uploaded every minute. To sort through all of this data and find specific recommendations for each viewer is why two neural networks are needed: one for candidate generation and one for ranking.
The candidate generation network sorts through billions of videos and provides broad personalization using collaborative filtering. This network takes events from the user’s history and retrieves a small subset of a hundred videos. Data such as IDs of video watches, search query tokens, and demographics are used.
The ranking network then has to filter through these hundreds of videos and rank them according to what the viewer is most likely to click on. It does this by assigning a score to each video using different features describing the video and the user. The highest scoring videos are then shown on recommended pages.
Even with this highly specialized system, YouTube receives criticism about it being a “misinformation engine” which radicalizes viewers by showing them conspiracy theories, fake news, and other disturbing content. YouTube keeps their algorithm close to their chest, so it is difficult to understand why this happens. However, it has become increasingly clear that disturbing videos are recommended more.
YouTube is constantly changing its model with new input from viewers and creators. In 2017, they supposedly began to improve the quality of videos by preventing inflammatory videos from popping up. In 2018, they added their controversial monetization policy, where clips can be eligible for making money depending on their content. This was meant to reduce the amount of content creators the platform had to actively monitor because YouTube has strict policies for what videos can get monetized. And yet, CNN reported that popular brands including Adidas, Cisco, and Hilton still had their ads running on extremist videos. This year, YouTube announced that it would be banning “borderline content” which could seriously harm or misinform viewers. The effects of this feature are still uncertain.
Essentially, YouTube uses an incredibly complicated “algorithm” which is made up of multiple components. Every YouTube video that you watch is delivered to you with a lot of metadata behind it. Now that’s something to think about the next time you scroll through your recommendations page.