Understanding YouTube’s Recommendation Algorithm for New Users
https://www.theatlantic.com/technology/archive/2018/11/how-youtubes-algorithm-really-works/575212/
Have you ever went on YouTube to watch a video a friend told you about, then once it ended a new suggested video popped up that caught your eye, and before you know it it’s an hour and a half later and you’ve been on your couch watching YouTube videos this whole time? I’m sure this is a situation that many people are familiar with, but I’m not so sure that people really understand what goes into which videos are recommended to them and catch them in this trap. Just like I’m not so sure people understand what goes into which webpages show up and in which order when they search something on Google.
My guess is most people would say it is based on popularity and personalization. Number of views is a huge factor in the algorithm, but there is a lot more to these algorithms than just popularity. This is because if recommended videos were solely based on number of views, then there would be an infinite loop that makes it nearly impossible for new videos to gain any attention. In this scenario, the most popular videos would consistently be recommended, which would result in them just gaining more and more attention, and this cycle would continue endlessly.
Furthermore, while personalization (another common guess) – recommending videos based on personal history of videos watched – is a main factor in this algorithm for consistent users, this is not helpful for a new user who has no personal history. For this new user, what factors go into this algorithm besides popularity?
As addressed in this article, YouTube designs its algorithm to adjust for this problem. In fact, this article points to research that shows that approximately 5 percent of recommended videos had fewer than 50,000 views. In order to filter through these new videos, YouTube learns from early performance of these videos. This means that they will channel through various new videos, recommending them to users for a short period of time, and then only continue to recommend them if they are performing well and show high potential.
This problem and solution is very similar to the problem in which we have an existing network of hubs and authorities, and we look at what happens if we add a new authority. If we think of the authorities as videos and the hubs as users, then if we were to add a new authority without connecting it to any hub, which is similar to adding a new video without recommending it to any users, it would gain no popularity even if it had the potential to be a good authority. However, if we could implement a learning algorithm that connects the authority to some hubs for a small period of time, and see if it gains value quickly, then we could decide whether it is a good authority or bad authority.