How will YouTube’s removal of the dislike count affect the user experience?
Recently, YouTube announced that it will be removing the dislike count in response to targeted dislike attacks against some media channels. Their hope is that it will reduce the number of these attacks and also relieve the mental stress of content creators. Now, people will only be able to press the dislike button, but cannot see the dislike count. In the homework, we were asked to evaluate whether or not we would stick more closely or less closely to the power rule when we got the option to receive reviews. Interestingly, the case with YouTube has become almost the opposite situation – information has now been removed, and how will that affect the spread of information.
We can have a look at how removing the dislike button affects the popularity of videos. For a user, the like-dislike ratio (number of likes vs number of dislikes) acts as an information cascade for the user – the user can treat the ratio as a signal, and given that signal, determine whether to watch the video, or to skip the video. By removing the dislike button, users are only able to see the number of likes on a video, so next best numerical metric of a video’s quality is how many likes a video got versus its number of views.
To examine how it impacts a user’s choice to watch a video or not, we can use Bayes’ Theorem to reason about the probability that a video is good or bad. Mathematically, define the event “G” be the event that a video is good, “B” be the event that the video is bad, S1 be the signal of the like-dislike ratio and S2 be the signal of the like-view ratio. We assume that a user watches a video if P(G|S1) > 0.5 or P(G|S2) > 0.5. The view-like ratio is less indicative of a video’s performance. In order to determine how good a video is based on the number of likes versus views, you would have to know the average number of likes for a given view count, then determine if the actual number of likes is higher or lower than the mean. It is difficult (and almost impossible) to approximate the true average like count of a video given a certain view count. Hence, the probability that a video is good or bad tend to be closer to each other. On the other hand, the like-dislike ratio is much more interpretable, so if a user sees a high like-dislike ratio, they are much more likely to watch a video. Conversely, if the see a poor like-dislike ratio, they are much less likely to watch a video. Hence, P(G|S1) > P(G|S2), but P(B|S2) > P(B|S1). This is the main concern of the viewers: since the like-view ratio is less interpretable, they are more likely to miss out on good videos and watch worse videos. We can also conclude that theoretically, removing the like-dislike ratio would make videos follow the power law less.
Luckily, there is another crucial piece of information that users can use to determine the quality of a video: the comments section. This is a place where people can leave their sentiment of a video, and determine the quality of a video. Hence, the effect of removing the dislike button is slightly mitigated by this. Furthermore, YouTube has moved to using a neural network to produce recommendations for its users. The network takes in features such as personal information (geographic background, age, sex etc.), language information (user’s search history, video content) and the video’s statistics (view count, percent of video watched, like-dislike ratio) to perform a multi class classification task. Thus the algorithm still has knowledge of the like-dislike ratio and many other features, so it will still recommend videos that you have enjoyed previously. While the removal of the dislike count is an inconvenience to many, in reality, the effect is offset by existing measures such as the comments section and the recommender system.
The removal of the dislike count, however, is indeed an effective method of preventing dislike attacks. We can think of dislike attacks as a collective action by a group B called “Bullies”. We can think of each person being an individual node with some threshold theta. In the beginning, every node has no edges to each other, except for when target groups form, in which case these clusters are fully connected. When a node B observes a dislike count, they are given information on how many people have joined the collective action, and can compare their threshold with the current count. If it is not high enough, B will not partake in the action. Otherwise, it will. In both cases, B will form a connection with those who have observed the dislike count on the video. Supposing that B would check on the video’s dislike count again after a while, B would be more likely to click dislike if the dislike count increased. If we removed the dislike count, however, no one can guarantee that someone else has disliked a video, so they themself would not want to dislike a video. We can thus see that removing the dislike count definitely helps when analysing from the perspective of graph theory and information cascades.
While YouTube’s decision to remove the dislike button may seem to be a big inconvenience to many, in reality, the effects are not too bad. On the other hand, the removal of the dislike count definitely helps reduce the number of dislike attacks. All of these claims fall in line with YouTube’s own analysis. Many other media platforms don’t have dislike buttons, such as Twitter. How will YouTube compare to these sites? Only time will tell.
Sources:
https://static.googleusercontent.com/media/research.google.com/en//pubs/archive/45530.pdf
https://blog.youtube/news-and-events/update-to-youtube/