Re:Mozilla Is Crowdsourcing Research into YouTube Recommenda...
by GuB-42 on Thursday September 17, 2020 @06:20PM
(#24510677)
I think these are exploitation/exploration trade-offs.
YouTube only knows your interests by the videos you watch. It knows you like speedruns, so it will show you speedruns, easy. But you certainly have other unrelated interests: cooking, woodworking, old cars, whatever. So, from time to time, Google takes a chance and shows you a random topic just to gauge your interest.
It works on a global scale too. They want some variety, they don't want to depend only on a few topics and previously successful YouTubers. So sometimes, they may pick a random video and so it to everyone, and if it succeeds, they have something new.
It is actually rather clear when you look at your recommendations to determine which are which. Typically you are going to find a mix of videos related to the video you just watched, videos relative to your global interests, trending videos, and random stuff.
If some generally effective machine learning algorithm acts really stupid from time to time, chances are that it is exploring, that is: it is asking you a question instead of giving you an answer.
Edit: someone mentionned the multi-armed bandit, yes, this is such an algorithm
YouTube only knows your interests by the videos you watch. It knows you like speedruns, so it will show you speedruns, easy. But you certainly have other unrelated interests: cooking, woodworking, old cars, whatever. So, from time to time, Google takes a chance and shows you a random topic just to gauge your interest.
It works on a global scale too. They want some variety, they don't want to depend only on a few topics and previously successful YouTubers. So sometimes, they may pick a random video and so it to everyone, and if it succeeds, they have something new.
It is actually rather clear when you look at your recommendations to determine which are which. Typically you are going to find a mix of videos related to the video you just watched, videos relative to your global interests, trending videos, and random stuff.
If some generally effective machine learning algorithm acts really stupid from time to time, chances are that it is exploring, that is: it is asking you a question instead of giving you an answer.
Edit: someone mentionned the multi-armed bandit, yes, this is such an algorithm