"the value of social media is shifting away from content hosting and removal, and towards recommendation algorithms directing one’s attention"
I really wish I could hack on the youtube recommendation algorithm. It seems to be tuned for zombie viewing pleasure, and totally not for finding surprising new content. Even just having several "personas" would be a big help.
Yes I feel like personal ownership of the recommendation engine is the key battleground..otherwise whatever the data is a third party is always deciding what you see.
Totally agree. The lack of configurability around recommendations is really frustrating. So often, the next video that plays is something tangentially related to what I'm watching, but still not what I want to watch.
I'd love a feature that defaults the next video to the newest unwatched video in a given channel or something like that, but we all know why it won't be built.
The problem is the revenue model. Zombie viewing is more profitable than surprising new content.
If you were asked to hack on YouTube's algorithm, any modifications you'd produce, no matter their merit, would be tested against the status quo for their ability to generate revenue.
I'm sure that is what they tell themselves at youtube HQ. But I'm not convinced. Why should one A.I. fit everyone? It just seems like laziness, ie. corporate inertia, to me. If they could open up an API for third party recommenders this would really be interesting.
Not really. Surprising new content has more active engagement which is crucial for acquisition ads. Revenue for pay-per-click is always better than display.
Yes. It creates another problem on a different level, which is that creators now try to hack the algorithm and generate contents that are more likely to be recommended.
Hm I disagree with that. I think for the short term it’s obviously aligned for zombie viewing but I’ve been introduced to completely different and new content via their recommendation algorithm on the homepage (no, not trending).
without user history to build a video similarity embedding with PCA from the user to video click matrix (or word2vec like click prediction), you would have to build something like .. topic based page rank.. and then show the highest ranked videos per topic.. that would require a fucking lot of mining
I wish they would made different algo for different "categories" of people, and you could just select from a dropdown list the one that works best for you. It would be win / win imo.
I really wish I could hack on the youtube recommendation algorithm. It seems to be tuned for zombie viewing pleasure, and totally not for finding surprising new content. Even just having several "personas" would be a big help.