|
|
|
|
|
by brudgers
2107 days ago
|
|
[semi-speculation presented as if I actually know] Youtube recommendations are best effort. That effort has a computational budget and a time limit. The fastest and least computationally intensive recommendation is to pull the last set of recommendations out of the on-device local cache. The next fastest thing is to recommend whatever is cached near the edge of your internet connection whether it fits your previous viewing or not. It generally does not do this. Instead it recommends edge-cached material with cosign similarity to your preferences. However, there might not be much of that if your preference is for outlying content. The slowest and most computationally intensive recommendation would be to compute a new tensor of your preferences and run cosign similarity across the entire vector space of Youtube. It doesn't do this for obvious reasons even though that's what we all want. Of course, this ignores the issue of Google revenue. Advertisers targeted you, target you based on the videos you've watched. So there's that too. Google recommends videos for which it has relevant advertisers. |
|
Google does not have that much interest in serving new, interesting videos over what earns the most clicks and therefore advertising. A lot of users might also be pretty fine with what is already served. Although I have seen more and more alternatives popping up, several with a monthly payment plan instead. While free content is nice, a paid plan incites different business practices than one where revenue is closely coupled with the number of clicks.