3 ms·

It's crazy to me how Google almost telepathically guesses what I will search for in all of the web, but it doesn't know what videos I will watch. YouTube's rec

It's crazy to me how Google almost telepathically guesses what I will search for in all of the web, but it doesn't know what videos I will watch. YouTube's recommendation algorithm is one of the worst I use. I have to actively manage my subscriptions and many people I know mass click "not interested" on videos to try to improve it. Spotify, Amazon, Netflix, and pretty much every other recommendation works like magic by now. There's something really wrong with it.

indigochill 8y ago
I had a hilarious incident of this with Google Search being on the money and YouTube being so far off the mark it appeared broken. Try searching for "GPL3" and I assume you'll see something similar. On the other hand, searching for "GPL v3" actually gets relevant YouTube results, so they are out there and YouTube knows how to find them if you ask the right way.
aylmao 8y ago
Google Search, Amazon, Netflix-- my guess is there's just have more structured data for these recommendation systems. Amazon and Netflix get high-quality descriptions and information tables about their videos/products, Google Search has to do a bit more digging, but it has the whole text corpus link structure, etc of a site. Spotify probably has it a bit tougher, but my guess is playlists play a big role as they are curated clustering of similar music (as opposed to just a "folder" to put all videos in a series, like they are on YouTube). Music itself is also probably much easier to analyze than video: extract the BPM and frequencies and you can tell a lot about the composition of type of song it's from. Video on the other hand-- you can do very computationally intensive CV, but that won't tell you much about the mood, meaning, genre, etc. of video something is. YouTube many times will have to go off a title, a very vague description and a ton of irrelevant comments. Edit: Actually, they have captions. I'd assume dialogue could give them a lot of relevant information, who knows ¯\_(ツ)_/¯

indigochill 8y ago

I had a hilarious incident of this with Google Search being on the money and YouTube being so far off the mark it appeared broken. Try searching for "GPL3" and I assume you'll see something similar. On the other hand, searching for "GPL v3" actually gets relevant YouTube results, so they are out there and YouTube knows how to find them if you ask the right way.

aylmao 8y ago

Google Search, Amazon, Netflix-- my guess is there's just have more structured data for these recommendation systems. Amazon and Netflix get high-quality descriptions and information tables about their videos/products, Google Search has to do a bit more digging, but it has the whole text corpus link structure, etc of a site. Spotify probably has it a bit tougher, but my guess is playlists play a big role as they are curated clustering of similar music (as opposed to just a "folder" to put all videos in a series, like they are on YouTube). Music itself is also probably much easier to analyze than video: extract the BPM and frequencies and you can tell a lot about the composition of type of song it's from. Video on the other hand-- you can do very computationally intensive CV, but that won't tell you much about the mood, meaning, genre, etc. of video something is. YouTube many times will have to go off a title, a very vague description and a ton of irrelevant comments. Edit: Actually, they have captions. I'd assume dialogue could give them a lot of relevant information, who knows ¯\_(ツ)_/¯