How Youtube Changed Its Algorithm
Nowadays, going on Youtube is a way to let the time pass, and procrastinate. However, it wasn’t always like that that. As Casey Newton from The Verge wrote on his article, “How Youtube Perfected the Feed”, Youtube started off as a platform to ‘embed videos onto other websites’. If we were to put this into terms of what we learned in class, Youtube first started off as more of a hub site — it offered a list of videos that could be use as a list of resources to a specific topic/site. However, as ideas such as the News Feed started to grow on the Internet, Youtube wasn’t able to keep up. Youtube’s way of individualizing what someone sees on their Youtube page was by having users subscribe to channels, which wasn’t doing much. Thus, when Cristos Goodrow became in charge of Youtube’ s recommendation system in 2011, there was a change from subscriptions to views. According to Business Insider’s Jillian D’Onfro, his article “The ‘terrifying’ moment in 2012 when Youtube change its entire philosophy”, Goodrow changed the recommendation system by using machine learning to recommend videos that were similar to other videos the user has seen.
At first, this change seemed to be the answer to Youtube’s problems. Youtube videos were getting a lot of clicks and views. However, users weren’t spending a lot of time on the videos they were watching. Instead, they would hope around from one video to another. The reason for this is because the program to find new videos for the user didn’t always find videos that the user liked. For example, a user would look for fight videos and would be recommended to videos with a thumbnail and title that looks like there was an actual fight, but the content itself would be of someone only talking about the fight. While the user is clicking through multiple videos, the algorithm would still see it as a video the user liked, since it was still clicked on, and thus would recommend similar videos, even though it isn’t what the user wanted. To put it in terms of our class, where a node is a video, Youtube started off by using a system where videos were linked together by similar titles and thumbnails, even if the content was very different. The score of whether the user wants to watch the video is based on what the user clicks on, and if a user clicks on a specific video, then the score of other videos with similar titles and thumbnails would also increase. However, since users were looking through videos they didn’t like, the score of those videos and similar videos still increase since there was a click. As a result, users would continuously be recommended videos that they didn’t necessarily like. This was problematic because although more videos are being clicked on, the time users were spending on Youtube wasn’t increasing much
To solve this problem, Goodrow changed their algorithm to only count a video that the user continued to watch and didn’t leave. In other words, using the same model as before, the score of a video would only increase if a video was watched to the end. This lead to a huge fall in views in March 2012, and as a result, there were also many angry creators who weren’t getting as many views as before. However, the time spent on Youtube itself increased, leading to more ads being viewed, and thus more money overall for Youtube.
Youtube continues to grow today through the use of Google Brain, Google’s artificial intelligence department. WIth Google Brain, new links between videos were created that software engineers before couldn’t have seen. The changes can be seen in new watching habits — now approximately 70% of the videos watched on Youtube are based on Youtube recommendations. With continuous small improvements overtime in the Youtube algorithm, Youtube eventually became what it is today. Although Youtube was big before, it continues to grow and become a big part of people’s everyday lives.
https://www.theverge.com/2017/8/30/16222850/youtube-google-brain-algorithm-video-recommendation-personalized-feed
https://www.businessinsider.com/youtube-watch-time-vs-views-2015-7